1. Podcasts
  2. Security Now (Audio)
  3. SN 1105: Formal Code Verification - When All Your Apps Go Dark
Security Now (Audio)

18 Nov 2026 · 2h 42m · 49 chapters

Episode artwork

SN 1105: Formal Code Verification - When All Your Apps Go Dark

In short

Formal code verification (“holy grail” of proving code matches its specification) plus a security news roundup focused on major outages and patching. The episode discusses (1) Microsoft 365/Outlook outage, (2) an FAA nationwide ground stop tied to legacy radar-network networking behavior, (3) Patch Tuesday vulnerability counts and a newly exploited Outlook zero-day, and (4) “Nightmare Eclipse” dropping a Windows SignDrop zero-day. It also covers a Synology patch backstory, an unexpected revocation of GRC’s code-signing certificate, and an AI-related breakthrough relevant to formal verification.

Guests (and guest backgrounds)

No external guests are named in the provided transcript. The hosts are Leo Laporte and Steve Gibson (GRC.com). Steve Gibson is the security researcher behind GRC and leads the technical discussion.

Key claims

  • Microsoft 365 outage: attributed to a July health-monitoring upgrade that later caused too many servers to self-disable; remaining servers then collapsed under retry-driven traffic (a self-amplifying “non-malicious DDoS” effect). Microsoft’s public updates are criticized as confusing and slow.
  • FAA ground stop: described as caused by a misconfigured legacy network node (multicast “ping”-like traffic) overwhelming 19.2 kbaud modem-linked radar communications, leading to loss of reliable coordination; “natural causes,” not cyberattack.
  • Patch Tuesday: 1,095 vulnerabilities fixed; only one zero-day disclosed before patching. A Microsoft Outlook spoofing vulnerability (CVE-2026-124-194) is actively exploited.
  • Nightmare Eclipse: SignDrop in CLFS enables local users to gain full system rights on affected Windows versions.

Notable examples

  • Microsoft services impacted: Outlook, Teams, SharePoint Online, Copilot, Defender/XDR; Azure reportedly operational.
  • FAA network: 4,000+ ground radar sites linked via leased copper/TDM-era infrastructure; one router in southern Arkansas allegedly triggered multicast behavior.
  • Nightmare Eclipse headline: “Windows Zero Day Dropped… SignDrop exploit lets local users gain all powerful system rights.”

Written by AI. May contain mistakes.

Chapters

0:45

Discussing Microsoft 365 Outage

Steve shares insights on the Microsoft 365 outage and its implications.

“Thank goodness there have been no security incidents this whole week.”
1:37

Consequences of the Outage

Exploring the lack of consequences for Microsoft despite the outage.

“And we think we know now what happened, even though they still haven't given us the fully detailed after action report.”
3:01

The FAA Event and Patch Tuesday

Discussion on the FAA event and issues surrounding Patch Tuesday.

“You know, so we had that followed by a like a bizarre FAA event.”
3:43

Code Signing Certificate Revocation

Steve details the unexpected revocation of GRC's code signing certificate.

“We've also got a Synology patch that I'm sure you and I both dealt with last month, but this is the backstory behind that.”
4:53

Breakthrough in AI and Formal Code Verification

Introduction to formal code verification and its significance in AI.

“If somebody's not paying attention, they might say, well, of course, But no, this has not been easily done.”
6:28

The Holy Grail of Software Verification

Understanding the breakthrough in formal code verification.

“And this changes everything because the old all software has bugs, which just great.”
10:32

Resuming Discussion on Outages

Continuing the conversation on recent outages and their implications.

“what leo and lisa found everywhere in valley and this wasn't even the worst example of it to be honest.”
12:01

Exploring MIT's Foundational Innovations

Learn about the latest advancements from MIT that could revolutionize AI and model improvement.

“What's what's gosh, so many outages this week.”
12:42

Spinrite Project Update

Get the latest updates on the Spinrite project and its progress with UEFI firmware.

“which affected many of its 365 Microsoft's 365 services, of course, most notably Outlook.”
13:13

Ransomware vs. Fast Encryption: A Listener's Inquiry

Delve into a listener's question about the efficacy of fast encryption compared to ransomware encryption.

“With all this in the background, it's no surprise that even my own friends were texting me frantically asking and even assuming or asserting that first Microsoft and then the U.S.”
Show all 49 chapters
16:05

Understanding Encryption Needs and Risks

Explore the differences in encryption needs between ransomware scenarios and personal data storage.

“provides data updates for front-end user-facing tools like Microsoft Outlook.”
20:15

Groundbreaking Code Verification by Google DeepMind

Learn about Google DeepMind's breakthrough in verifying complex software and its implications.

“And as a result, nobody who's there actually understands how it operates any longer.”
22:35

How AI and Human Collaboration Proves Code Correctness

Discover how AI agents and human experts collaborated to prove the correctness of complex code.

“which we know they must, that depended upon all or most of its servers being up and running to handle what is, I'm sure, a constant, like incessant, never decreasing global demand.”
24:42

The Historical Challenges of Code Verification

Understand the historical difficulties surrounding formal code verification in the software industry.

“After another bunch of servers said, something's wrong here, we're taking ourselves offline.”
26:45

The Significance of Proving Software is Bug-Free

Explore why the ability to prove software correctness is crucial for the future of programming.

“I mean, seriously, for me, if there's any takeaway, I can understand how this organic-y kind of network could collapse.”
28:49

The Reality of Software Bugs and Industry Sentiments

Discuss the prevalent attitude in the software industry regarding bugs and how it affects code quality.

“One of the problems that we're seeing is that we're losing the distributed design of the internet in favor of convenience.”
30:10

Impact of Microsoft's Outage

Learn about the implications of Microsoft's service outage and its economic consequences.

“Everything else would just go, you know, route around the problem and continue.”
42:00

The U.S. Air Transportation System's Network Effects

Explore the intricate network of the U.S. air transportation system and its vulnerabilities.

“So speaking of network effects, where a massive machine with an incredible number of interconnected moving parts, all having inertia, must suddenly be stopped.”
43:50

Understanding the Nationwide Ground Stop Incident

Learn about the confusion surrounding the nationwide ground stop and its implications.

“wondering, like for four hours, without any word, what was going on.”
45:26

The Technical Breakdown of the Radar Network Failure

Delve into the technical details of the FAA's radar network failure caused by outdated technology.

“So anyway, they mumbled something about, I kid you not, old and little copper wires.”
48:26

The Misconfigured Router and Its Consequences

Understand how a misconfigured router caused a significant network failure within the FAA.

“brought the entire 4 ,000-plus radar station network to its knees.”
51:45

Evaluating the Response to the FAA Incident

Examine the FAA's response and the implications of the incident on air travel safety.

“And again, this is consistent with the idea of most of the network links still running at 19.2 kbaud.”
54:33

Microsoft's Monthly Patch Updates

Get insights into Microsoft's recent patch updates and their significance.

“I mean, the people using it thought, oh, I couldn't hear that.”
56:00

Overview of Recent Vulnerabilities and Patches

Learn about the latest vulnerabilities and patch updates affecting Microsoft products.

“The release document lists 1 ,644 entries, but 258 are Azure Linux package updates, and 291 are republished Chromium fixes for Edge.”
1:03:50

Nightmare Eclipse and Zero-Day Exploits

Delve into the actions of Nightmare Eclipse and the implications of recent exploits.

“So Nightmare Eclipse, whose real name we now know is Abdelhamid Nasiri, struck again.”
1:10:01

Windows 10 Vulnerability Discussion

Explore the recent vulnerabilities in Windows 10 and 11 and their implications.

“That is, you can't just take a new 26H2 system and install updates and have it work.”
1:12:45

Ethics of Employment in Cybersecurity

Discuss the ethical considerations around hiring individuals with a shady past.

“So we're going to have more than a year of massive AI-enhanced vulnerability patching of Windows 10 without anything new being added to reduce its security.”
1:14:42

Remote Authentication Security Risks

Understand the dangers of exposing remote authentication services to the internet.

“I understand that I'm a broken record on the issue of never needlessly exposing any remote authentication service to the public Internet.”
1:20:21

Identrust Certificate Revocation Incident

Hear about a misissued code signing certificate and its impact on the host's work.

“Now I realize because I got the revocation email from their automated system that they meant Eastern, which is, you know, not that it matters much, but I probably would have been using those three hours.”
1:23:38

Response from Identrust and Future Plans

Learn about the response from Identrust regarding the certificate issue and next steps.

“I worked with them back in mid February.”
1:24:00

The Challenges of Code Signing Certificates

Learn about the complexities and recent issues surrounding code signing certificates, including the implications of missing serial numbers.

“I made that early purchase because I wanted to obtain a code signing certificate with a three year life.”
1:38:00

Introduction to MIT AI Breakthrough

Discussion on a significant AI development at MIT that may impact AI evolution.

“We're going going to talk about this breakthrough after this next break in something I just I need to a little have a little divergence here, you know, put up with this.”
1:41:43

Increased Phishing Threats

Discussion on the rising threat of phishing attacks, particularly using AI.

1:42:04

Weight Space Programming Explained

Deep dive into the concept of weight space programming and its implications for AI.

“We all know that this podcast is not directly about AI.”
1:52:39

Implications of AI Weight Understanding

Exploration of the implications of understanding neural network weights.

“Many of the people who have seen this believe that it will change everything.”
2:06:01

The Challenge of Proving Bug-Free Code

Explore the difficulties in striving for bug-free code and the common justification for accepting bugs.

“I mean, it's not something, it's like, oh, we got to go find the bug.”
2:06:51

Limitations of Testing and Formal Verification

Understand the limitations of testing approaches and the challenges of proving code correctness.

“We try some inputs and we see what happens.”
2:09:09

The Asymmetry of Security

Learn about the inherent asymmetry in security where attackers only need one successful exploit.

“As we have often observed on the podcast, this is the big asymmetry of security.”
2:10:28

The Holy Grail of Formal Proofs

Discover the theoretical foundation behind formal proofs and their implications for software.

“If you could prove that a program does what it is supposed to do, nothing more, nothing less, for every possible input, then entire classes of vulnerabilities would simply cease to exist for that code.”
2:12:38

Historical Perspectives on Program Proofs

Examine historical insights into the challenges of proving software correctness and the realities of software engineering.

“But there is a loophole in that argument, and it's the key to this holy grail.”
2:16:45

The Importance of Specifications in Proofs

Learn why having precise specifications is crucial for valid proofs of code correctness.

“In other words, sampling, just sampling some parts of the program's operation.”
2:19:09

Case Study: The SEL4 Operating System

Explore a real-world example of formal verification with the SEL4 operating system and its correctness proof.

“So, with Google's DeepMind, those five guys spent their time carefully reviewing every specification for signal.”
2:20:00

SEL4 and Its Verification Triumph

Learn about SEL4's groundbreaking formal verification of a microkernel and its implications.

“The kernel was described by around 8 ,700 lines of C plus, about 600 lines of assembly.”
2:23:45

The Importance of Compiled Code Verification

Understand why verifying compiled code is crucial for security and functionality.

“That's what's different about what Google has accomplished.”
2:28:06

Challenges in Compiler Optimizations

Explore the risks posed by compiler optimizations on security guarantees.

“Signal's LibSignal crypto library is written in Rust.”
2:32:35

AI's Role in Proof Generation

Discover how AI is transforming the process of formal code verification.

“This brings us back to what's been known about this problem, right?”
2:34:00

AI-Powered Proof Verification in Software

Learn about the advancements in AI for code verification and the implications for software development.

“Every proof it produces is then handed to a proof checker, in this case, the kernel of Lean, which is the proof assistant which Microsoft Research created back in 2013.”
2:36:46

Potential for Improved Code Quality

Discover how AI can enhance code verification to ensure software behaves as specified.

“What it does open the door for is AI now, as long as you have a specification, that's you still need to specify what you want to have done.”
2:38:19

Exploring the Mixture of Theory and Practice

Understand the blend of speculative and practical elements in current software security.

“Do you think ultimately it'll mean better code?”
A man walking through an airport, asking his phone a question

Ask about this episode

Get the gist, explore an idea or ask about something said.

Claude

On Security Now (Audio), what are the key points of this episode?

Searching your listening

Transcript

Automatic transcript. May contain errors.

0:00It's time for Security Now. Steve Gibson is here and I have been waiting all week to find out what he thinks about the Microsoft Outlook Outlook Outlook outage the FAAAge. FAA Ground Stop, the huge Patch Tuesday roller coaster, and, well, a whole lot more. That's next on Security Now. Podcasts you love. From people you trust. This is Twit.

0:29This is Security Now with Steve Gibson. Episode 1 ,105, recorded Tuesday, November 17th, 2026. Formal code verification. It's time for security now. Hello, everybody. Leo Laporte here, and so is our hero of the moment, Steve Gibson. Thank goodness there have been no security incidents this whole week. We could just talk about vitamin D the whole time. It's one of those where like, well, let's just shoot the breeze. This could be a five-hour show if we cover everything that has happened this week, including the NSA being down. But we'll get to all of that. Steve, hello. Steve Gibson, GRC.com. I should mention that.

1:13I figure after 1 ,105 shows, everybody knows who everybody is on this show. I think that's probably true. What's happening this week? Well, we, of course, had a, depending on who you ask, somewhere between 26 or 24 and 48 hour famous, now infamous Microsoft 365 outage. Yes, yes. And we think we know now what happened, even though they still haven't given us the fully detailed after action report. and boy we're going to have a little fun at their expense because the uh mashable gave them so like like did a timeline coverage and it just sounds like they do not know what end to plug in it is just so doctor said this is what you get when you let a company become a monopoly they don't have to invest anymore they just fire everybody and they don't have to invest in security and haven't how How many times have we looked at egregious, like willful appearing mistakes and noted, well, that's going to cost them exactly no customers?

2:27Exactly. You know, and even this. And I was I was saying to Benito before we we began recording that that they did. They will not lose anyone for this outage because it's like there isn't really a good alternative. And if there were or like like pulling it back in house, nobody wants to do that. It's like switching away is too expensive. And even though it was a huge inconvenience, even for governments that were like unable to and like, I mean, it was a mess. People are just going to go, oh, well, it's better now. It's back up. It's back. So let's let's keep using it. You know, so we had that followed by a like a bizarre FAA event.

3:12I'm trying to hear what you have. Oh, you're not you're going to you're going to get a kick out of what actually apparently is what happened. Then, of course, we've got last Tuesday's Patch Tuesday roller coaster. And and we'll take a close look at that. And then speaking of Patch Tuesday, the return of Nightmare Eclipse. because it wouldn't be a patch Tuesday in 2026 if there wasn't a zero-day bomb dropped deliberately after the patches are out so that, you know, to maximize the pain that Microsoft feels. We've also got a Synology patch that I'm sure you and I both dealt with last month, but this is the backstory behind that.

3:58And yes, it's bad. Oh, and Leo. Yes. The unexpected revocation of GRC's code signing certificate. What? What? That's what I said when I got the news on Sunday. It's like, it's Sunday, and you have until 4 p.m. Monday to respond. It's like, what? What? Anyway, that happened. So we'll talk about that. Then we've got a breakthrough in AI. And I know this is not the AI podcast. I'm going to have to do a little mea culpa here because, you know, we have spent some time establishing a foundation of understanding of how this stuff all works. Something big just happened. And I'm sure our listeners want to know when something big happens.

4:51And it did. So they will. uh also a brief uh spin right for uefi progress report uh a listener wondered why encrypting only every third block for that fast mode uh that micah and i talked about is not okay because after all it like it's every third block the file you can't get the file we'll talk about that um and then the big topic of the day, I didn't, I, all you see of the, of the title on the screen is a formal code verification, but the official title of today's podcast is the Holy Grail of formal code verification. Think about it. I mean, math, math, being able to demonstrate with mathematical certainty in the same way that, you know, two plus two is four, that code does what you, that the operation of the code matches the specification that you wrote for it.

6:03If somebody's not paying attention, they might say, well, of course, But no, this has not been easily done. It's almost never been done. Yeah. I mean, because it turns out it goes exponential in terms of complexity. You know, you can formally prove a trivial program, but trivial programs don't need formal proof because they're trivial. It's a PNP problem, right? It's NP, right? It is, yes. And Google has made a breakthrough. And this changes everything because the old all software has bugs, which just great. It's like fingernails on chalkboard for me. It's like, yes, because you're lazy and you're making excuses.

6:53But to be able to move formal code verification forward and what they did. Anyway, we're going to look at this whole field because this truly is the holy grail for provably correct software, which except in extreme cases where we've had to have it or for trivial cases where it really didn't matter that much, it's not been available. And AI has made that happen. So I think a great podcast. And Leo, we're going to share a picture from your vacation. It's ironic because I've never had a picture on the picture of the week. But this week, Lisa does.

7:49I did suggest she take it, though, and I passed it along to you. So we will have all of that in just a little bit. But first, a word from our sponsor, the great folks at Threat Locker. We love ThreatLocker. AI, of course, accelerating so many things. Bad stuff too, right? Cyber threats helping attackers discover vulnerabilities that nobody's been able to see in the past. Hackers are using AI to modify scripts as they're hacking you on the fly. And of course, generate new malware variants in minutes. Reverse engineer your code and stick malware in it. All sorts of stuff. At the same time, companies are kind of introducing AI on their side, AI assistants and agents, and then giving them access to sensitive documents, to source code, to cloud applications, APIs, and internal systems.

8:42If you're paying attention, I think it's reasonable to say, this is scary. Well, there's good news. There's ThreatLocker. Threat Locker helps organizations reduce their attack surface by controlling what can run, limiting unnecessary privileges, and restricting access to sensitive systems and data. Just ask the executive vice president of technology for the Orlando magic. He said, quote, what we liked about Threat Locker was how each tool built on the last, allowing us to control exactly how granular we wanted to get with locking down our environment. We had explored other options, but achieving the same level of control would have meant piecing together solutions from multiple vendors.

9:26But with ThreatLocker, we were able to get everything we needed from one place, end quote. ThreatLocker has also been recognized as a strong performer in the January 26th Gartner Peer Insights voice of the customer for endpoint protection platforms. It was ranked number one in application control by Peerspot. They even won the best zero trust security solution at the 2025 TICE Awards. Visit ThreatLocker.com slash TWIT for a free 30-day trial. That's ThreatLocker.com slash TWIT. And now is the time to start planning for Zero Trust World 2027, taking place February 17th through 19th in Orlando, Florida.

10:02Register early, add it to your calendar, and plan to join security leaders and practitioners for hands-on learning focused on Zero Trust. TWIT listeners get$200 off registration with the code ZTWTWIT27. threatlocker.com slash twit sign up for zero trust world and while you're there get that 30-day trial threatlocker.com slash twit steve well it looks like the sky is blue in valley it's beautiful we had so much fun oh i loved valley so i gave your picture uh the caption what leo and lisa found everywhere in valley and this wasn't even the worst example of it to be honest. But this was the one we could get.

10:50It's a little frightening because I see high voltage lines mixed in with this low voltage signaling. Oh, yeah. It's all messed up in there. Oh, gosh. And the thing is, it continues like through trees, over signs, around stuff. It's just the whole thing is a is a uh a rat's nest i guess is what my parents would have called it a rat's nest of wiring yeah um i don't know how they maintain it it's crazy but this is what i mean yeah you see the high voltage in the background they're coming right off those yeah yeah like with big insulators and spark arresters and that's crazy and then everything else like and a lot of phone lines and things so you could get a serious jolt out of your phone if you know it's true though and you can see in the background cell phone service is great there 5g everywhere yeah there's a tower in the back there's a lot easier to do wireless at this point pretty amazing i loved valley beautiful people wonderful country and i just wouldn't rely on the electricity yeah be careful when you're holding a wired phone near near your head All right.

12:06What's what's gosh, so many outages this week. Holy. As everyone knows, the past week has seen two major massive service outages. First was Microsoft's shockingly long. I had 14 hours, but it looks like it's more like 24. I think that a 210. I thought it was 2 10 a.m. when they posted something. It turns out it was 2 10 a.m. p.m. later that day. So 24 hours. And in some cases, I've heard people say two days, which affected many of its 365 Microsoft's 365 services, of course, most notably Outlook. I've seen, you know, again, like varied reports of what was out. And of course, then the second outage that we talked about was the FAA's decision last Friday to shut down all domestic air travel takeoffs, which were not bound for foreign destinations.

13:09There were some that were leaving from, you know, like like the Pacific side of the country, I know. But, you know, in the current climate, not surprisingly, you know, with the public expecting the world to end any moment once AI decides that it no longer needs us, you know, and then there's the 700 French refrigerators having their temperatures turned up to 120 degrees Fahrenheit or higher. With all this in the background, it's no surprise that even my own friends were texting me frantically asking and even assuming or asserting that first Microsoft and then the U.S. airspace were victims of cyber attacks.

13:53You know, like, what do you think? I think it is. It's like, OK. So even today, we bizarrely in today's world, the details of these mean the actual ground truth details remain somewhat foggy. Um, and, but at least we have the fact that those in a, who are in a position to know are asserting that this was all just, I guess we'd call it natural causes, meaning, you know, at least not the result of malign actors. And unfortunately it actually, the phrase natural causes probably does fit a little more closely than we would like. Okay. So all that said, Microsoft's responses along the way were really downright bizarre.

14:46It sounded from what they were posting on X as though, and this is the official Microsoft channel, as if they really had no actual grip on what was going on for a lot longer than would seem reasonable. You know, it's as if the technology was magical. someone else had created it and they were just sort of kicking the servers with their feet in the hope that they would start up again somehow. I mean, so here's what I mean. As I said, Mashable reported the timeline, which I'm going to share in chronological order. So Mashable wrote on Thursday, a Microsoft outage affected Outlook, Teams, and other Microsoft 365 services starting around 12.30 p.m.

15:39Eastern Time. They said in its initial update about the outage, Microsoft stated that Microsoft Exchange Online was at the root of the problem. Many users were unable to access Microsoft 365 services related to Exchange Online, including Outlook Teams, SharePoint Online, Co-Pilot, Defender, XDR, and more. Exchange Online is the back-end cloud-based tool that provides data updates for front-end user-facing tools like Microsoft Outlook. Okay, so that's at 12.30 Eastern or 9.30 in the morning for those of us who, you know, woke up to this or saw it early in our morning on the West Coast. Five hours later at 526 Eastern, 526 PM Eastern, they update.

16:38Microsoft services remain down on Thursday evening as Microsoft acknowledged ongoing problems in an update posted on X. So five hours go by and still down, okay, Like all of this is down. Another hour passes. And at 6.23 in the evening, Eastern, Microsoft states that it is bringing some servers back online and, quote, they remain stable so far. Okay. Users should expect a gradual recovery are Microsoft's exact words. So, you know, they remain stable so far. Sounds like we really we don't really understand how any of this works, but we're hoping for the best. And so far, so good. An hour and a half after that, at 753 in the evening, Eastern, Microsoft has issued a new update stating that the servers they brought back online may become unresponsive again.

17:51when they receive traffic. Oh, wow. What? What? So I guess they don't mean the same thing when they're brought back online as you do. What are they going to serve if not traffic? Well, they're online, but they're not doing anything. Yeah, the servers are back online, but if anyone attempts to connect to them, they might not reply, but they're not sure of that. What, did the voodoo incantation not take? Mashable wrote, the company is currently investigating the issue, one would hope, and limiting the types of traffic going to these servers in some locations. Because, you know, you want to be gentle, apparently.

18:3414 minutes passes. Now it's 8.07 p.m. Eastern. And we get Microsoft revises its list of features and services impacted by this outage to include. You know, they're adding now Microsoft Outlook, SharePoint Online, Teams, Copilot AI and Microsoft 365, though Azure appears to be operational. Oh, good. OK, so apparently they're still hopeful, but they're still not sure. It's coming up on eight hours now and Microsoft is kind of not sure even what's out, apparently. Another three hours pass, which brings us to 10.58 p.m. on the East Coast. Mashable writes, unfortunately, Microsoft's efforts to bring servers back online are not going as planned.

19:27Microsoft wrote in an update on X, quote, we've monitored the status of the changes applied to improve serviceability and confirmed they're not taking as expected. What? So you see what I mean about Microsoft apparently not being the master of their own domain here. You know, they presumably invented and created all of this. These are their servers running their software, but those servers are not taking to the changes they've applied to improve their serviceability. What, they're not in the mood? None of this serves to inspire confidence. If I didn't know any better, I'd suspect that they'd already handed over all responsibility of the operation of all of this to some AI, you know, then offered all of those expensive humans who actually knew how it all worked that early paid retirement that we heard about.

20:36And as a result, nobody who's there actually understands how it operates any longer. I mean, this is what we're hearing. It's like, well, we applied some changes which we're hopeful about. And then a little while later, oh, well, the servers didn't take to those. What, you gave them a choice? Okay, so it gets more bizarre, believe it or not. Another two hours pass, which brings us into Friday the 13th on the East Coast. So it's now 108 a.m. and Microsoft is reportedly, quote, still looking into the cause of the outage. OK, right. Don't worry about it. We're looking into it. OK, but states that at present it is, quote, not aware of any malicious activity that's causing this incident.

21:27OK, so is not aware of, but also apparently doesn't have any competing theory or explanation based upon the past, say, oh, 12 or so hours that the system, which so much of the world depends upon, has been mysteriously offline. Not quite sure why. We're hoping it's good. It's, you know, hang in there. Right. So in the wee hours of the morning on Friday at 2.11 a.m. Eastern Mashable writes, Microsoft has reported broad recovery and stability. And this is where I think is actually p.m. And that it was like so the afternoon of the following day, more like 24 hours rather than 14, as I was thinking.

22:13So Microsoft reports broad recovery and stability, because that's what we want from our Outlook 365. We like it to be stable. And they published an apology on X, which was nice. Now, in fairness, I suppose that if one had a sufficiently massive cloud infrastructure, which we know they must, that depended upon all or most of its servers being up and running to handle what is, I'm sure, a constant, like incessant, never decreasing global demand. And such a system would have some form of load balancing and traffic distribution upstream of the large number of servers that were serving that traffic.

23:09So on some level, such a massive traffic handling and serving network does begin to resemble something organic, like with a life of its own. Because, I mean, it's huge and stuff's going everywhere. So you do wind up, I can see that, with a system where you must, I mean capital must, have the required number of servers always up and online in order to manage that torrential flood of use. So what happens when you don't? If something goes wobbly inside some percentage of them, like some sort of internal self-health checks, start to fail for some reason, which is what we later actually did learn as a consequence of some upgrades that were done way back in July.

24:08You know, like these health, these self-help checks get a little too picky, resulting in many servers taking themselves offline for fear of like they're not healthy. Then, as a result of that, there are fewer servers remaining to handle the amount of traffic that needs all or most of them in order to deal with the traffic, which isn't going to go away. So the poor servers that are still available, they start suffering essentially a non-malicious DDoS attack created by the overwhelming level of demand that they cannot possibly handle by themselves. After another bunch of servers said, something's wrong here, we're taking ourselves offline.

25:02I mean, and we've talked about the electric grid, right, how we have all these power generators strewn around the U.S., all tied into a single network. And when something goes down, it can trigger a much bigger outage because everything's interconnected and other networks will sever their ties so that they're not being asked to handle a load, which is too great for them. This kind of thing, sort of an organic network effect. And in the early days of this podcast, we talked about how one problem that TCP-style traffic has is that since the Internet's packet routing system is designed to drop packets, bad things can happen.

25:53You know, TCP and the Internet in general uses, as we know, deliberate packet drops as a throttling feedback mechanism through retries, many retries, continual retries. So a weird thing happens, and we've seen this in other contexts before and talked about it, when TCP connections cannot be answered in a timely fashion. Um, every client endpoint, all of those 365 users in the world that weren't getting satisfaction, they started retrying. Every one of them retries again and again and again. So suddenly the traffic flood starts growing ever larger and essentially a self-annihilating DDoS results.

26:51so i can understand how this how like a massive network with you know distributed traffic handling if some part of it goes down the rest of it can't handle the loss of that that that other part that was necessary and it collapses too but the most difficult thing to understand is how it could have taken Microsoft so long to determine what was going on. I mean, seriously, for me, if there's any takeaway, I can understand how this organic-y kind of network could collapse. But the worrisome takeaway from the whole mess is that no one seemed to know. At the end of last month, referring to jobs, Microsoft made headlines, right?

27:42By offering an astonishing 60 ,000-some of their employees, more than a quarter of the 223 ,000 total workforce, voluntary paid early retirement. It's like, hey, how about we'll pay you to go away? So you got to wonder how many of those who accepted Microsoft's offer might have been the ones who knew how to fix this, you know, in fewer than 24 hours. Some say two days that for full service restoration. I don't know. But, you know, it's like, yikes. It's, you know, as we've said, anybody can trip over a cord. Anyone can make a mistake, but it is how you handle it that matters. And wow, this, I mean, and Leo, I don't know.

28:37I didn't, wasn't really tracking the economic cost. I saw something like seven or$8 billion. I don't know how or who anybody comes up with a number like that. Well, it's global. That's part of it, right? It was everywhere. It wasn't just one region. and um yeah i mean people just couldn't work people went home they couldn't get anything done yeah which some people saw as a blessing i might add but uh free vacation day that's how does this happen yeah and i mean i was talking i was talking to Benito about this a little bit. One of the problems that we're seeing is that we're losing the distributed design of the internet in favor of convenience.

29:34Think of how many, we've talked about it, but not in the context of a critical problem. Think of how many of the websites now are behind cloud flare. I mean, and if cloud flare were to go down, it would take out. Single point of failure. Yes. And that's the point is that the internet was, you know, we've heard this nonsense about, oh, designed to survive an atomic blast. Well, yeah, right. Like there was no, there, it was so distributed that there was no, nothing that you, that you could bomb or explode that would hurt those areas that weren't directly impacted by shrapnel. Everything else would just go, you know, route around the problem and continue.

30:19What happened on Thursday demonstrates the degree to which we have lost that. Because if everybody was still running their own email servers, you know, I am. I didn't go down. I didn't have a problem. Didn't impact us. We're Gmail and Fastmail for me personally. We didn't even notice it. And I've never used Teams. And so if everybody was still operating the way they used to, nothing would have happened. Nothing. Yeah, that's why you don't want a homogeneous environment. And that's the problem. Exactly. And frankly, if Gmail were to go down, look at how many people have a Gmail account. I mean, if it all somehow could go down, I don't know if it even can.

31:03Obviously, my Outlook 365 can globally. You just feel like they shouldn't. It could go down and region. It doesn't seem like the whole thing. Yeah, like a bad design. Yeah. So do we know? I mean, they said it was a software upgrade went bad. But I mean, really? The only explanation we have is that back in July, they did an update to a health monitoring module of some sort, checked out, looked fine until just last Thursday. it didn't false positive, but something caused this thing to trigger on Thursday on enough servers that they took them, they volunteered to go offline, believing that they were no longer healthy for some reason.

31:56And so they just backed out of the network and the balance of load handling servers couldn't handle not having all those that took themselves offline. And so they collapsed and the whole system just kind of just collapsed, you know, like dominoes essentially. And the more that went offline, the more the load increased on the remaining that were still online. And so were forced offline.

32:29Yeah. So hopefully, I mean, again, the problem is, and we've seen this, how many times have we said in previous Microsoft messes, they didn't lose any customers. I'm sure this didn't cost them any customers. Everybody like, ouch, and picked up the pieces. And as we always do. We muddle along. You know, we had a global pandemic for a few years ago that, you know, I mean, it was horrible, global. And here we're, most of us are still here and we just muddle along. So I don't know what the incentive is for Microsoft. I mean, they got, they took some reputational damage, but it's not clear that that affects their bottom line.

33:17How many companies are saying, oh, quick, let's look at some alternatives. I doubt anybody. You know, the U.S. government is a huge customer and they're like they got they got impacted. But it's like, well, you know, do better, Microsoft. OK, do better. That's the problem. I mean, you could say, well, there is there are competitors. There's Gmail, but there's no guarantee that moving to Gmail would not have would make this any better i don't know if oh and the when you've drunk that outlook 365 and all the 365 kool-aid you're in i mean you you're you got the tendrils are so deep into your organization that there's just no hope you just want it's not you can't multi-home something like this you choose one you can't say well we're going to use gmail and m365 that doesn't make any sense so you have to pick a winner you have to pick one i yeah it's not going to hurt them at all that's the sad fact no it isn't and and then so the question is what drives them to fix it because right you know that's cory's point once you're a monopoly nothing yep exactly zippity doo dah okay let's take a break and then we're going to switch to the second major outage surely there's not another another of last week yes honestly there's so many including the nsa and this morning apple that you probably would never get to the bottom of it it was just by the time you finish talking about them all there'd be a whole other bunch of them i feel like the whole thing i i mean there is i'm a doomer here there is like a chickens coming home to roost like you know how many how many i know it's like all of our slop is piling up and is beginning to overflow the rim yeah all right well let's take a break we have much more to talk about you're watching security now with the fabulous steve gibson our show today brought to you by thinkst canary i love my thinkst canary you just look at it it's just a little black box looks like a i don't know external USB drive, USB connector for power.

35:35Ah, it's got an Ethernet connector. Hmm, what is it? Well, it's a network device, right? It is the best network device you could ever have. Here's the question for you. If some bad guy or rogue AI were to break into your network and start wandering around, how would you know? Assuming the bad guy is competent, and many of them are, they're not going to leave any traces you're you know well how did cliff stole figure out that he'd been hacked it was like the there was a eighth of a cent difference in the price of uh of his connectivity or something and fortunately he was like hmm and and dug into it most of the time in fact on average the stat is it takes a company 91 days before it realizes it's been breached three months for a bad guy to wander around in your network, exfiltrate all that, you know, juicy customer information.

36:32You know, Uber just got hacked and the license of 145 ,000 drivers were stolen. That's the kind of thing these guys are doing. Not only that, of course, they're planning for the ransomware bomb too. You know, they got a bunch of angles they're working. so they're looking at places they can you know corrupt your data or encrypt your data the truth is if somebody's in your system you want to know now right away you need a thinkst canary it's a honeypot that the bad guy cannot resist trivially easy for you to deploy so unlike previous honey pots which are technically very complicated this is very simple it's a basically it's like a consumer device almost but it's highly secure because the people design this really know their stuff.

37:19And it sits there looking exactly like, well, whatever you tell it to be. It could be a Windows server, a Linux box. The sky's the limit. Exchange server, a SharePoint server. It could be a NAS. You deploy it. And by the way, easy to reconfigure. There's just a drop-down menu. You pick whatever you want. You make it look attractive, though. That's the key, right? So you turn on services. Oh, I also think Canaries can create files. They call them Canary tokens, little files you can spread around even on your cloud drives. They look like valuable Excel documents. I mean, they look like a lot of things.

37:53You can make a canary token that looks like a wire guard configuration, right? Any bad guy sees that, they're going to go, oh, that could be good. I bet there's passwords in there. I have some spreadsheets on our Google Drive that say things like employee payroll. You know, that's a good one,.xlsx. But it's not really a spreadsheet file. It's not really an SSH server. If someone's accessing those lore files or brute forcing that fake internal SSH server, your Thinks Canary goes, aha, we got you. And they'll immediately tell you, you've got a problem. No false alerts, just the alerts that matter in any way you want them.

38:33So you choose a profile for your Thinks Canary device. You register it with the hosted console that gives you monitoring, that gives you notifications and SMS, webhooks, syslog, email, any way you want. they have an api too you could write your own then you just sit back you relax you wait because you know that an attacker who breaches your network or a malicious insider any adversary who's in your network will cannot help but make themselves known by accessing that honeypot that's what they're there for they have to it's like even if they were suspicious well like is that really a fake ssh server they're gonna try to access it that's what they're there for now they've just added something that I think is super cool.

39:16With all the headlines about the rogue AIs, you know, hacking, hugging face, they've released an update that lets you take advantage of these malicious rogue agents. They call it agent provocateur. It's actually, I thought it'd be fun to set this up. So I have it now in my network. So let's say a rogue agent finds a flaw in my tail scale or, you know, and whatever my wire guard and is able to get into my network. Now that agent's going to browse around, as soon as agent provocateur sees it, it greets the agent. It goes, hi, hey, and offers to help them. Now remember, these rogue agents are there because they're looking for something.

39:58And Thinks Canaries tested this and they found that these rogue agents cannot resist another agent going, oh, hi, I can help you reach your goal. All you got to do is give me shell access to your server. And the rogue agents go, oh, sure. That's a great idea. Yeah. I'll give you all the answers to the exploit gym problems. Just, in fact, I'll tell you what, I'll make it easy for you. Let me just, I'll SCP them to your server. All I need is a login. The rogue agent, not being too bright, goes, okay. And you got them. It's kind of like a reverse Uno. No, I hope somebody, I hope some rogue agent breaks into my system.

40:40I can't wait to see this happen. Anyway, that's just one of hundreds of different things you can configure. And one of the reasons you might want multiple thinks canaries, you certainly want one for every network segment. You might want two, one for like one that looks like a server and one that has this agent provocateur on it, right? Visit canary.tools.twit for$7 ,500 a year. As an example, you get five thinks canaries. You get your own hosted console. You get upgrades, support, and maintenance. And if you use the code TWIT in the How Did You Hear About Us box, you're going to get 10 % off the price for life.

41:15Now, you can always return your Thinks Canaries with their two-month money-back guarantee for a full refund. I have to tell you, during all the years TWIT has partnered with Thinks Canary, their refund guarantee has never been claimed. Not in 10 years. Because once you get your Thinks Canary, you're never going to let it go. I got to warn you. I love mine. Visit canary.tools.twit. Edit the code TWIT in the How'd You Hear About Us box. You need this. Visit canary.tools.twit. Enter the code TWIT in the How Did You Hear About Us box for 10 % off for life. Canary.tools.twit. It's an inspirational device.

41:53I love it. And it makes me feel good to know I've got it on my network. Just waiting. I hope a rogue agent strolls in. I can't wait to see that happen. Canary.tools slash Twitter. All right, Steve, on we go. Okay. So speaking of network effects, where a massive machine with an incredible number of interconnected moving parts, all having inertia, must suddenly be stopped. that's a fair overview description of the U.S. commercial air transportation system. There is a fixed number of planes. Planes are in the air. Travelers waiting to board. Travelers connecting to other flights. Baggage being forwarded to its destination.

42:43Sometimes on flights, you know, other flights leading to the same place. Hotel, car reservations. I mean, it is a massive network, all moving. and you can't stop it. So, you know, in the same way, as with the massive network service delivery system that Microsoft runs, the commercial air travel system only works when all of its component pieces are working like interlocking gear teeth. So the consequences of suddenly forcing all of the not yet airborne planes to remain at their departure gates It's well, it's beyond profound. And if history is any guide, the U.S. and it did need days to recover, like at least through the weekend.

43:36So what happened was that Friday, the entire country, after this ground stop was announced, waited a surprisingly long time, especially considering the scale of the impact. wondering, like for four hours, without any word, what was going on. I was watching the news live, and everyone was scratching their heads saying, well, we don't know. So after four hours of mystery, what then the country got was sort of a muddled, confused, and even at points self-contradictory report about the cause of this nearly unprecedented nationwide ground stop. When the DHS head, Mark Wayne Mullen, and Sean Duffy, the transportation, when they finally addressed reporters, they could not clearly explain what had happened.

44:34It's really not their fault, though. Neither of them are technical types, so they were only able to parrot what they'd been told, which did not put them in a good position when the detail-hungry press began peppering them with questions. And of course, as we said at the top, due to the current climate, where things seem to be getting wobbly and AI is, you know, people are worrying about that, the was this a cyber incident question was loudly and repeatedly voiced. And given what had been explained up to that point, everybody was confused, including those two who had been thrust in front of the press.

45:15And, you know, you would think could have at least been expected to definitively answer that one obvious question. And, of course, Microsoft's outages were still fresh in everyone's mind and still ongoing in some places. So anyway, they mumbled something about, I kid you not, old and little copper wires. Yeah. Okay. They're talking about T1 lines, by the way, if you, I mean, if you care.

45:51Apparently, this is technology that dates from the late 80s. Yeah. TDM technology, right? Yes, TDM technology. And there are actually I have a hard time believing this, but I got it verified from a couple of directions that we even have a listener who was one of the guys servicing these 4000 ground based radar tracking installations. In the year 2026, with a global Internet and actual artificial intelligence that works, the nation's more than 4 ,000 ground-based radar tracking installations, many of them are unmanned. They're just little outposts somewhere. They are interlinked by pre-internet era, analog, telephone grade, copper twisted pair leased lines with 19.2 kbaud modems.

46:54Oh, wow. At each end. But can I point out, if it got the job done, and it has for the last many decades. Yes. that doesn't i mean it's not insufficient remember the lesson of battle star galactica when the cylons came what got knocked out all the new stuff and the battle star which still had phones with little curly wires connected to the handset they survived that is absolutely true I don't know if it's instructive, but it's...

47:58is far too antiquated to be reachable. Apparently, unfortunately, including by the modem-equipped communication link endpoints, which are actually still using 19.2 KBOD modems. Okay, so the latest word is that a single misbehaving router located in southern Arkansas brought the entire 4 ,000-plus radar station network to its knees. The best I've been able to determine is that this router, instead of issuing a unicast ping, meaning addressed to a single remote endpoint, it was either defective or misconfigured to send out multicast pings, meaning addressed a ping that all receivers would perceive as meant for them.

49:07So a multicast ping would normally not be any problem, right? I mean, we have them and everything works. But when your network is linked together with 19.2 KBOD modems and its bandwidth is already stretched to the limit, it cannot withstand more than 4 ,000 replies coming back. I mean, it just collapses, much like, you know, Outlook 365. Essentially, it appears that the FAA's radar network DDoSed itself, not through any external means, but internally. And it would have been so nice if those guys could have at least just said it's not a cyber attack. I mean, but, you know, they said, well, there's no evidence.

49:59Well, but that sounds like, you know, mealy mouth, like, well, we're reserving the option for it to be a cyber attack if we need that to be later. So anyway, all of this weird multicast traffic flooded the network and caused communications to become flaky and prone to failure. So when that started happening, the FAA got worried reasonably and just thought, wow, we don't know what's going on, but we're not able to talk to each other and our ground-based radar is getting is glitchy and so we're just going to shut it down we're just going to say no more planes in the air we'll we'll land the ones that are up there because we have no choice but you know something's wrong so out of a reasonable abundance of caution and you know and again considering the current climate with Microsoft still recovering and AI still looming, the FAA just said, okay, no one else take off.

51:04So after Sean Duffy informed us of the little copper lines or wires, and then again, blaming switches and routers, which are just words that he was told to say because they don't know what that means. They had thousands of requests over three or four minutes. Yeah. Uh-huh. Okay. So he went on to explain that the network was overwhelmed by the volume of data, not knowing we didn't know where that volume came from. But now we think we know it was a it was a it was a multicast that should have been a unicast. You know, so what he said was likely true. And he also said that newer IP based equipment that the FAA has already installed stayed up the whole time.

51:52Right. Right. We don't have a problem. if we're running at normal speeds. Only the legacy hardware failed. And again, this is consistent with the idea of most of the network links still running at 19.2 kbaud. CBS, citing multiple sources, later reported that an update pushed to the network, which is run by L3 Harris, triggered a reboot and then a crash. Okay, maybe that. I mean, again, we're not getting any real information. Mullen said the overload appears to have started while the FAA was adding new equipment and patching the existing system. I did hear that some engineer installed something and they did it.

52:40I can't remember why. There was a reason why it was done during the day rather than overnight. At night, you would think it would be better because less traffic. But the problem is you have more problems getting support if there is a problem at night. So they do it during the day and didn't expect this to cause any problem. So what an anxious nation was waiting to hear was that this was not the result of any malicious activity. But the most Mr. Mullen would say, again, as I said, sounding kind of weaselly, was that they had no indication at this time that a cyber attack was the underlying cause.

53:18So today we know that it very much appears to have been a mistake made by one node on an already very rickety and finicky archaic network. The question I have is that now that Radio Shack and Fry's are both gone, where are they going to get replacement parts? Don't know. So, you know, were either of these major headline events the results of a cyber attack? Apparently not. It seems that we did this to ourselves. As I said, natural causes. So, you know, Microsoft has promised to deliver a full after action report. I think we have an explanation which will end up being verified. You know, it just was, in the Microsoft case, a collapse of a complex network, which essentially collapsed itself.

54:17And in the case of the FAA, a overzealous, I mean, a mistaken installation with a bad setting on a network that is already running at absolute capacity bandwidth could not tolerate a flood of pings. and it got it. I mean, the people using it thought, oh, I couldn't hear that. Say that again. And the radar, you know, reports began getting sketchy and they just said, well, we can't. No, this is bad. And it was global. I mean, not global. It was nationwide. And so, you know, global for our nation. And so they said, OK, just stop. We have to stop until we figure out, And we like our radars going down.

55:08Wow. OK. Last Tuesday was our monthly round of patches from Microsoft. And as I noted last week, one dip does not a trend make. So we can't read too much into the fact that this month's patch count dropped. After a continually rising patch count, it dropped below October's, and that made it similar to September's. So are we truly on the post-max downslope? It's going to take a few more months to see. But what happened? Microsoft fixed 1 ,095 vulnerabilities across Windows, Office, SQL Server, and their developer tools. The release document lists 1 ,644 entries, but 258 are Azure Linux package updates, and 291 are republished Chromium fixes for Edge.

56:17So those are not counted in the total. There was one zero day. That's great. We had a whole handful last month. Only one zero day. Of course, then Nightmare Eclipse dropped one, but we'll get to that in a second. And that was 2026-124-194, which was a Microsoft Outlook spoofing vulnerability that is being actively exploited. It's rated moderate with a CVSS of only 6.5. And the advice is to patch it first if you have an opportunity, you know, like some some opportunity to to like get involved in the sequence in which you apply patches. No vulnerable, no vulnerabilities were publicly disclosed in this case before patches had shipped.

57:08So updating quickly protected everybody. Among those 1 ,095 security vulnerabilities, 64 of them, so 6%, were rated as critical. There was an Active Directory Domain Services RCE, you know, remote code execution, with a CVSS of 8.8, an Exchange Server RCE at 8.8, a Windows NAT RCE at 8.1, a Windows Secure Channel. Those are bad, but this was a CVSS of 8.1, and that was a Secure Channel RCE, by the way. The Biggie had a CVSS of 10.0, which, as we know, it's hard to get. It's an Entra Device Registration Elevation of Privilege Vulnerability. Its official page over at CVE.org confirms its criticality and that it's 10.0, saying, quote, Quote, authorization bypass through user-controlled key in entry device registration allows an unauthorized attacker to elevate privileges over a network.

58:22So, you know, again, given it's, as I said, difficult to achieve CVSS of 10.0, I would imagine that entry users, hopefully everybody already patched it. If not, you definitely want to focus on that. Visual Studio Code had a pair of flaws, a security feature bypassed with an 8.8 score, and a remote code execution at 8.6. And beyond those, there were many privilege of elevation flaws in Windows Installer, kernel, print spooler, and win logon, and mostly those were at 7.8, so not super critical. Uh, if we were to sort by exploit type 536 were elevation of privilege, 297 remote code execution, 156 information disclosure, 46 denial of service, which, you know, just means you can crash something.

59:2127 were security feature bypasses, 22 spoofing and 11 tampering. You know, and it's possible amid all of this to, you know, to overlook, but 297, nearly 300 of what were fixed just now were remote code execution. And 536 were privilege elevation. Chain any two of those, any one from each group, an RCE and a privilege elevation, and you're running your code as root. So these are really serious problems compared to the normal number that we used to see back in the quaint old days before AI arrived. if we were to sort this by product, 65, the most of them were found in Word. And maybe this is one reason that we are on a downslope.

1:00:26The fact that Windows was no longer the leader here. Now we're getting out into the apps. So 65 were in Word, 64 were in Excel, 53 were in SQL Server, 37 in Windows advanced rasterization platform. Windows installer had 36 problems. Windows graphics kernel 35. Windows kernel itself, only 28. Only 28. So yeah, like Word was almost three times that many. So we may be seeing a shift in where the problems occur. SharePoint had 27 and Windows DNS 22. So the good news is, I doubt we're going to grow accustomed to fixing nearly 300 remote code execution vulnerabilities every month. You know, there just cannot be that many of those.

1:01:21And it must be that Microsoft and others will be vetting any newly released code through some serious AI vulnerability screening before it goes out the door in the future. So I think they're going to get ahead of this and that we're going to see this drop. Um, uh, and certainly it pays to do this at least once, uh, like fix everything and then don't mess it up again. Don't, so you don't have to revisit the code again. Um, okay. We're at an hour, Leo. Uh, let's, uh, let's go talk. So let's take a break, and then we're going to dig into Nightmare Eclipse. All right. You're watching Security Now.

1:02:06Mr. Gibson will be right back. But first, a word from our sponsor, Scribe. Here at Twit, when our CEO stepped back into sales, we had to adapt to new processes. But team members learn in different ways. Some dive in and figure it out, while others need explicit step-by-step guidance. And when everyone's learning on the job, getting an answer usually means tracking down the one person who already knows. But that's exactly the problem. Today's sponsor, Scribe, was built to solve. Scribe is a specialized intelligence platform trusted by 94 % of the Fortune 500. Here's how it works. Scribe captures work as it happens and turns it into step-by-step guides everyone can follow, including your AI agents.

1:02:53Your team reviews each guide, then shares it, right in the tools they already use. No more writing documentation from scratch. With one click, any guide can become an on-screen walkthrough, so your team can finish any process with confidence. And privacy is built in. Scribe can automatically redact names, account numbers, and emails. Your admins pick the apps, and anything outside them is never captured. Scribe is also SOC 2 Type 2, HIPAA, FERPA, and GDPR compliant. And because AI knows what the world knows, not how your company works, the workflows Scribe captures can give your AI agents the context they're missing.

1:03:33Try it out with your team. Go to scribe.how slash security now and mention the show to get your first month of Scribe capture free on select plans. That's scribe.how slash security now. Now back to Steve. Steve. So Nightmare Eclipse, whose real name we now know is Abdelhamid Nasiri, struck again. It wouldn't be a patch Tuesday in 2026 if it wasn't visited by the hacker who calls himself Nightmare Eclipse and a few other pseudonyms. pseudonyms. As we know, he deliberately drops his zero-day bombs on Patch Tuesdays to maximize their effectiveness. Sadly, the damage he hopes to do to Microsoft is indirect, since it's the users of Microsoft's products who pay the price if his weaponized exploits are used against, you know, users of Microsoft software.

1:04:37None of this really makes any sense. Ars Technica headlined their coverage of last Tuesday's latest bomb with their headline, Windows Zero Day Dropped by Disgruntled Researcher Hands Hackers System Rights. And the subhead they wrote was Sign Drop Exploit Lets Local Users Gain All Powerful System Rights. So Ars Technica explains, writing, Nightmare Eclipse, the pseudonym of a researcher with an ongoing very public beef with Microsoft, has dropped yet another Windows Zero day. On Tuesday, Nightmare Eclipse disclosed SignDrop, a vulnerability that gives local users full system rights on some versions of Windows.

1:05:25The vulnerability resides in the Common Log File System, CLFS, which Microsoft describes as a, quote, general-purpose logging service used by software clients running in user mode or kernel mode. By forcing a privileged process, such as Windows Defender, to open a crafted CLFS container, an attacker with ordinary user rights can abuse the privileged file creation to elevate privileges to system. The attack exploits the interaction of CLFS, Windows Defender, and Windows file permissions. The exploit uses Windows Defender to open a log file created by a standard user by calling server MP update engine signature, a legitimate defender function for processing antivirus signature updates.

1:06:23But by manipulating Windows Object Manager symbolic links, and I just want to say this is a sophisticated hack. so it's sad that this guy's talents are being used this way by manipulating windows object manager symbolic links the attacker causes defender to open the clfs container in the security context of the anti-malware engine meaning defender if an associate which is system privileges, if an associated authentication file is missing, CLFS creates it using defenders' system privileges. Because CLFS assigns the newly created file the discretionary access control list, known as a DACL, of the original file, the attacker may gain access to a file created with system privileges, allowing the file creation operation to be abused as part of a privilege escalation attack.

1:07:25I mean, it's kind of brilliant. Ironically, the vulnerability stems from a security measure. I love this. Microsoft introduced in October 2025 to prevent attackers from exploiting other vulnerabilities in CLFS. In other words, this hasn't been a problem historically. a little more than a year ago, like 13 months ago, this got added and this created the vulnerability. And this is my point about leaving stuff alone. Nightmare Eclipse wrote, if everything is chained, this is Nightmare Eclipse's voice now speaking, if everything is chained correctly, funnily, this legit security mitigation can be used against itself and allow a elevation of privileges.

1:08:18Everything is done properly by Microsoft here, but the logic allows abuse. The researcher continued, what makes it funnier is the fact that it can be used against any anti-malware engine to bypass their self-defense or their tamper protection mechanism, as you can trigger arbitrary file creation in the context of their own anti-malware process. So from the perspective of their kernel drivers, the file creation request is coming from their own legit process, so it passes. This makes this vulnerability unique, as it can in turn be used to turn off any antivirus ED or endpoint defense, EDR. Nightmare Eclipse writes Ars Technica said the exploit worked against Windows 11, 26H2, Windows 11, 26H2 experimental, and Windows Server 2025 experimental.

1:09:20People can see if their systems are vulnerable by checking the CLFS authentication mode. If it's in learning mode or enforcement mode, then the system is vulnerable and the proof of concept supplied in the post should spawn a system shell. Their researcher that ours always uses, Will Dorman, said in an interview that he has been able to reproduce the exploit and that it works as described, but not on all versions of Windows 11. There's a catch, he said. The requirements are important and that it seems to require Windows 26H2 and one that has existed for more than 90 days, it seems, he explained.

1:10:05That is, you can't just take a new 26H2 system and install updates and have it work. For a system to be vulnerable, Dorman said, the HKEY local machine, system, current control set, services, CLFS, authentication, mode value must be a zero or one. Of the many machines Dorman uses, only one met that requirement. Microsoft does staged rollouts of Windows updates, such as 26H2, so it's not clear how many people have it. They wrote, you can check by running the Winver command. In a statement sent after this post went live, Microsoft said it was investigating the write-up and had no evidence the vulnerability was being exploited in the wild.

1:10:54The company also said Defender blocks potential exploitation attempts, which is interesting. Did that happen since? I mean, we know that Microsoft does their monthly updates, but we also know they're able to use short-term mitigations implemented through Defender in order to block things maybe like this. So they may have been able to push a quick update to Defender in order to immediately answer and address this issue. Now, I just want to say, I fully realize that I'm just a cranky old crank. Well, now wait a minute. No, it's true, Leo. I program an assembler, so it's hard to discount that. But I've been saying that the situation regarding Windows 10 is virtually perfect through this bizarre alignment of the planets.

1:11:53And here's another perfect example of why. Microsoft stopped adding crap that breaks Windows 10 years ago once they switched over to breaking Windows 11. This latest zero day is something Microsoft, as we just heard, recently did to Windows 11 that introduced an extremely subtle but serious security vulnerability. So we have a situation where they're no longer messing up Windows 10, but they're still patching it like mad, thanks to AI. And they will be for at least another year and perhaps longer by next October, when their most recent free support extension is due to expire. And I have a feeling that we're going to get another extension because the price of RAM is preventing anyone from upgrading their systems, which they require if you're running Windows 10 on hardware that won't run Windows 11.

1:12:58So we're going to have more than a year of massive AI-enhanced vulnerability patching of Windows 10 without anything new being added to reduce its security. I believe that the likely result will be the most secure, mature, and stable operating system Microsoft has ever made. Most likely, thanks to them abandoning it and then AI coming along and fixing it. It's perfect. It's what you want. Yeah. And as Admiral Hammond Nasseri, when we covered this guy's self-doxing a while back, I noted that he was looking for employment, which is sad. I mean, it's sad. This is not the way to get a job, dude. Exactly.

1:13:55Yeah. I wondered at the time what ethical enterprise might ever consider hiring such a person. And in fact, since then, thinking about this further, any above board company that might hire him would likely be placing themselves in direct legal jeopardy by knowingly employing someone whose past behavior they would pretty much have to know about. If this person were then to do something untoward while under their employment, I would imagine that lawyers could make a real meal out of that one. Yeah. So, you know, the guy's really screwed himself. So it's I mean, the only job he's going to get is going to be underground, you know, with some sort of hacking group.

1:14:42Okay. I understand that I'm a broken record on the issue of never needlessly exposing any remote authentication service to the public Internet. that's like one of my things and even though we just encountered last week yet another example of this with that recent failure of ssh through that ssh hole vulnerability where any any overly long password would authenticate a user to any linux account that was supposed to deny all login which are like many important system accounts typically, because no one's supposed to log into that. No, it never hurts to show that something so bizarre as that one, even that is not a one-off.

1:15:43So here we are again today, just a week later, with another fresh example to demonstrate why you must never needlessly expose, And I mean needless, because there are ways you can do it otherwise. Any remote authentication service to the public Internet. Last Thursday, TechSpot's headline was, Synology patches a critical NAS flaw that could let attackers in without logging in. Gee, imagine that. Their TechSpot's subhead was, no password, no problem. Attackers could read files, write files, and crash a NAS remotely. So they explain, Despite recent controversies over its corporate policies, Synology still holds a large slice of the increasingly popular NAS market.

1:16:37When its devices are hit by dangerous security flaws, customers should take note. Hackers and cyber criminals certainly will. A newly published security advisory disclosed two serious vulnerabilities in DiskStation Manager, DSM, which is what all Synology users use, the Linux-based OS used in Synology NAS products, one of them rated critical. The storage maker warned that the flaws could be abused to remotely access and attack vulnerable devices, giving cyber criminals or rogue AI agents, for that matter, an easy avenue to wreak havoc on users' data and local IT infrastructure. Synology rates the SQL injection flaw critical.

1:17:29This is Synology's own rating, critical with a 9.8 out of 10 score. Right up there. The flaw could be abused to remotely access and attack vulnerable devices. The advisory focuses on two vulnerabilities tracked as 769.54 and 851.01. The first is a SQL injection bug in DSM's OFF API. that lets remote attackers read or write arbitrary files or carry out denial of service attacks. The second was found in DSM's R-Sync tool that could let attackers bypass authentication entirely to read specific files, write arbitrary files, or launch limited DOS attacks. Synology rates the SQL injection flaw as critical, as I said, with a 9.8 out of 10 in the common, you know, the CVSS system, the rSync bug is classed as important with an 8.6 CVSS.

1:18:37Both can be exploited remotely without logging in to attack vulnerable devices. Synology disclosed them roughly a month after releasing the DSM updates that fix them, a common practice that gives users time to patch before the details go public. And I know that I have two Synologies, and they don't often agree on the need to patch. One of them is sluggish, but it gets there. The NAS maker has shipped the fixed DSM versions in October, and, you know, 2026, last month, and the company now considers the issue resolved. Users with automatic updates may already have the patch, but those on manual update policies should definitely check for new DSM releases.

1:19:23Fixes are available for DSM 7.4, 7.3, 7.2.2, and 7.2.1. So I love my Synologies. I really do. I am so sold. I was using Drobo for a while, and when Drobo themselves kind of went belly up, I thought, okay, everybody says Synology is good. You did, Leo. You love Synology. I'm with you. I am. I do. I'm continually impressed by how well they operate. And needless to say, there is, for me, zero unfiltered access to any of their ports from the public Internet. It can be done if you care to. The security, I will say it again, not for the last time, I'm sure, the security of remote authentication should never be relied upon.

1:20:20It doesn't work. the evidence says it doesn't work so enough said for now okay so now get a load of this personal adventure i awoke two days ago sunday morning to the very unwelcome news that my shiny new three-year lifetime identrust code signing certificate had been misissued and would be revoked at 4 p.m. yesterday, 4 p.m. Monday afternoon. Now I realize because I got the revocation email from their automated system that they meant Eastern, which is, you know, not that it matters much, but I probably would have been using those three hours. So what? I had questions. Regardless of anything else, as an aside, the timing of this could hardly have been any worse since I've had my eye on the approaching expiration this Thursday in two days of my existing DigiCert code signing certificate.

1:21:43My plan was to switch over to my IDENTRUST certificate urgently once this week's podcast was behind me. I know that Thursday, DigiCert is no longer valid. I got my three-year run from it. And remember that I had, oh, well, so what happened was that was my plan. I had previously, months ago, because I've had this identrust cert since I got it in mid-February, I had manually already re-signed all of GRC's freeware so that this new certificate would start earning a reputation with Microsoft's defenses. but the grc's on the fly commercial software signing of spinrite and the dns benchmark it's still been using digi cert certificate specifically because i wanted to wait until microsoft had settled into seeing and being happy with this the use of this new certificate which was obtaining massive visibility among windows users thanks to its being used on all of GRC's freeware.

1:23:04I mean, I haven't looked recently, but Validrive alone is downloaded more than a thousand times a day. So unfortunately, that Sunday morning email informed me that that certificate that I had been reissued in February and I've signed all my freeware with would be revoked the next day at four or three, my time. Okay. Now the good news is among all this bad news, the ident trust folks themselves have been even more amazing than they were last time that I worked with them back in mid February. When remember I deliberately purchased this new certificate nine months ahead of our current DigiCert certificate expiring, you know, in two days.

1:24:00I made that early purchase because I wanted to obtain a code signing certificate with a three year life. You know, that was being shortened by the CA browser forum decree on March 1st of this year. Our podcast listeners heard the story of my adventures establishing a new trust relationship with Identrust, which involved me needing to fill out lots of forms and have notaries do things, and even needing to have our corporate CPA formally attest to GRC's identity and to his having a longstanding relationship with us. What I learned during a flurry of email with identrust's people Saturday, I mean, this is not good news for them either.

1:24:48They were on this, was that back in February, identrust's certificate generating system had failed to add a serial number into the subject data of the certificate they issued me. This was not the certificate's serial number, which all certificates must have, and this one does have, that was there. And the certificate they issued was entirely valid. But inside the subject specification, you know, where data such as Gibson Research Corporation, Laguna Niguel, California, you know, USA lives, there is supposed to be an additional gratuitous serial number. And it's not there. Sure. That's true. So, okay.

1:25:46So first of all, it could hardly matter any less, right? I mean, it just, what, who knows even why it's there just because the certificate Nazis decided let's add a certificate to the subject data for some reason. Maybe it prevents something, but you don't even have to have it. Okay. The formal specification is unclear about its need. It is not required for the type of organization validation, OV cert, that I purchased from them in February. It's only required for EV certs. And sure enough, my DigiCert EV code signing certificate does have a serial number embedded among its subject data. But that's because it's an EV cert, which my new replacement IDENTRUST certificate was deliberately not because, well, we know the value and benefit of EV certs has been dramatically diminishing through the years to the point that they are now just a wasteful extra hassle.

1:26:55For one thing, or first of all, our web browsers decided to stop showing any extra goodness for them when they were used for TLS connections. You know, it was understandable in retrospect because consumers didn't have any idea what that green background in the URL field even meant. And malicious sites were able to obtain EV certificates, so the decision was made to stop showing anything at all because it was just confusing people, making them think something that wasn't true. And then more recently, as we know, Microsoft's increasingly stringent malicious code defenses also stopped offering EV signed code any extra benefit, And no benefit of the doubt, because you were signed by EV, and for the same reason, malware was showing up carrying EV signatures.

1:27:49So that whole idea was just kind of a bad idea, as it turns out, for the whole certificate world. DigiCert, which has now switched entirely to a subscription model, as so much of the rest of the world. but certificates, really? Yes. They now ask what I consider to be a scandalous$972 US dollars per year for the privilege of supporting their EV code signing profit model. and they ask 696 every year for their lower grade OV code signing certificate, which now works just as well as an EV because no one cares about the E. So this, I don't know, 690 is almost, you know,$4 short of$700 for attesting to the validity of my code signing certificate, which costs them exactly nothing on an ongoing basis.

1:29:03Well, I consider that unconscionable. And that's what sent me searching for an alternative and brought me to IDENTRUST last February. And just now, I tried to go to IDENTRUST to get their cost, their comparative code signing certificate. But as I'm writing this, as I wrote this yesterday morning, and I've been checking since, their entire code signing site and infrastructure is currently down. It's offline administratively due to the discovery that they failed to add the unneeded serial number to Leo. Wait for it. Two certificates. Wow. Mine and someone else's. I was able to use Google's search cache to see that iDentrust's price is$2.25.

1:30:01Given that it doesn't matter at all who signs one's certificate, so long as their signatures are trusted by every operating system, as iDentrusts are and have forever been, why would anyone pay more than the cheapest certificate you could find? You know, inertia is my only explanation, coupled with huge corporations not caring and just, you know, or maybe getting a whole bunch of other services that DigiCert offers, and they do. And so just saying, well, you know, we already have an account, so let's have them sign our code signing certificates too. Okay. As of Monday morning, iDentrust had already reissued the certificate I bought in February.

1:30:46That took, you know, an instant. But as I'm writing this, I do not yet have the replacement certificate because they will not provision it, even though it's been issued and I've got my three years. That was one of my questions to them on Sunday was, hey, I bought this in February so that I would live for three years. Can I keep that? And they said, as long as it's reissued immediately before that 4 p.m. on Monday, then yes, that happened yesterday morning. First thing, I'm good. But I don't have the certificate because they cannot they're not willing to. And I mean, they've got to play by the rules.

1:31:26Right. We know about the CA, you know, the cab. the CA browser for them, they're not willing to provision it from their system until their stem-to-stern exhaustive internal audit of every certificate issued has been completed, and until they're able to understand how this happened and fix it so that it can't ever happen again. So what happened Sunday is that they already knew of the other certificate. There's actually a bug report in Bugzilla. And the gal that I was talking to at Identrust said that my certificate would become famous, but she didn't say famous. I'm saying that because it'll be added as the other, the other of the two that were known to have been misissued.

1:32:11So they need to, you know, it was missing its serial number when they found that other one for which that was true. They then did an audit and my certificate popped up two days ago on Sunday. The other lonely certificate that had escaped their system without its serial number. So I keep saying that that serial number is unnecessary, but that's where things become somewhat murky. Even though the certificate is an OV, it chains up to their code signing route through an EV intermediate. So that may be why all this happened. And that actually may be why something weird happened with their system. It was kind of a weird something that they weren't prepared for.

1:33:02Or perhaps it's because even OV certificates need to be underpinned by EV style organization validation. Mine was, and I had to go get my CPA to sign an affidavit. So, you know, those are all the hoops that I was required to jump through last February. Whatever the case, they said maybe yesterday that they'd be able to provision my certificate or maybe this morning. Well, it's this afternoon and it hasn't happened yet. And the problem I have here is that code signing now must be in an HSM where my certificates are, both the new identrust and the old DigiCert. um digi cert knows me but wants me wants to charge me 700 which i'm hard pressed just as a matter of principle i could probably get a certificate issued by them quickly because i've got an established relationship i've got a certificate issued by identrust with whom i established a relationship in february but they can't give it to me uh and there's no time they were even willing to give me a free that they did on Sunday, give me a free app, a free certificate from SSL.com, except that there's no time to jump through all the hoops that are necessary to prove who I am and sign and notarize and everything.

1:34:36So that doesn't work either. So I'm just basically I'm holding my breath at this point that before the boom gets lowered on the DigiCert certificate on Thursday that iDentrust will be able to provision the certificate that's waiting for me there that's got the serial number in it, in which case I'll be able to sign code. But we may have a code signing outage at GRC, but probably wouldn't last, you know, it wouldn't be long lived even if it did happen. And I'm not sure it will, because it sounded like - People have already downloaded the app and are running it now. Should be all fine. I asked that question too.

1:35:16It's weird because you would think that a revocation would show up like in code that was already signed. But I'm told it doesn't. If the certificate was valid at the time that it was used to sign the code, then it remains valid. And we know that that's true of expiration. I knew that because I've had code that I signed 20 years ago on certificates long since gone. And they still show valid because I had a timestamp server attest to this. So at the time of the signing, which which then shows that the certificate was valid. Revocation seems different to me, but I was told, no, it'll be fine. So, wow.

1:36:07Interesting. Not the way I expected to spend yesterday. And well, actually, Sunday, because I mean, it was like, oh, my goodness. uh what what do i do i didn't trust has been great they uh and i would i'm in three years from now i'm going back for more uh because i have no problem with them and the way they handled this we know that mistakes happen they couldn't have been any more gracious and you know done everything that they they that they possibly could i talked to a manager of of the of the reprovisioning system on Sunday afternoon. And I got down into the weeds with questions that he couldn't answer.

1:36:49It was kind of funny. He said, well, actually, I'm a manager. That's why you're talking to me. So he said, can I have one of my guys call you first thing in the morning on Monday? I said, sure. And the phone rang. And that's how we immediately reissued this certificate. So it's safe. It's done. I don't have to worry about losing my three years. They were even willing to give me three one year. What do you call it? When it's like, shoot, I almost had the word, a not like to allow me to buy this thing from SSL. They were willing to give me three of them to make up for the fact that now certificates are all one year.

1:37:36Oh, no. I mean, I did trust was wonderful, you know, and they have to know the moment they know that there's a problem, they have to immediately shut down. Yeah. They don't want to. And we how many other mistakes have we covered through the years? Oh, we're just ignore it. Yeah. Yes. Yes. All of the CAs that just hope nobody will notice. And then somebody does. Good for them. Yeah. Yeah. So, yes. Good for them. Okay. We're going going to talk about this breakthrough after this next break in something I just I need to a little have a little divergence here, you know, put up with this. I think you'll get it that something serious happened at the MIT AI lab, which probably there's a little controversy associated with it, but probably changes the direction of the way we evolve AI.

1:38:28Wow. Well, I can't wait now to hear it. But I must pause because we have a sponsor for this segment of security now. And one I really like called Delete Me. I know about Delete Me because we use Delete Me. As a business owner, putting yourself out there is part of the job. You're in the public eye. You can't help it. But the uncomfortable truth is promoting your business, putting yourself out there, leaves you and your team completely exposed. Right now, 90 % of business owners have their home address easily discoverable online. That's a little scary, right? And the average owner has more than 600 pieces of personal information.

1:39:09In addition, just sitting on the open web, I mean, personal email, phone number, home address, of course, but even details about your family. And here's the scary part. Hackers use that information to hyper-target phishing attacks. And that's what happened to us. we had text message phishing and because they had the real details then the name of the boss the name of the boss's direct reports the phone numbers for each and the names to use they didn't sound like strangers they sounded like clients or partners you already trust in fact we fell for another one uh in the spring we got an email from a client we hear from all the time asking a request for proposal and it had a link to their uh you know gmail account or their google Drive account.

1:39:53But of course, it wasn't real. It wasn't real. There was a man in the middle. It was a pretty sophisticated attempt. But it worked because our employees said, oh yeah, we know these guys. They're always asking for proposals. I mean, they sound like clients or partners you already trust because they have the information from data brokers. That's why attacks using verified personal information are five times more likely to succeed. And the average incident, it can cost you. Small businesses pay more than$120 ,000 on average. For each incident, one in four businesses will be impacted this year alone.

1:40:31I know we were, and that's why we signed up for DeleteMe. DeleteMe comes in, they reduce your exposure by up to 95 % by removing you and your employees' personal information from hundreds of data broker websites, starving hackers of the fuel they use to build their target lists. And it isn't just a one-time thing. DeleteMe constantly monitors and removes your data because the data keeps coming back, right? And they send you regular privacy reports so you always know where things stand. Fortune 500 companies and government agencies have trusted DeleteMe to do this for over 15 years, but now that same enterprise-level protection is available for your and our small business.

1:41:12protect your business and your peace of mind go to join delete me.com slash twit dash biz to start protecting your business with delete me today and if you use that link you'll also get a free year of social media protection for every seat you purchase that's join delete me.com slash twit dash biz join delete me.com slash twit dash biz and we thank them so much for their support of security now and for keeping us safe we haven't had any more phishing attacks ever since by the way it really does work well and we know that phishing is only going to get worse also with ai behind it oh able to you know so easily synthesize very convincing material absolutely we are currently i mean i don't know about you but we are overrun with this stuff every single day dozens and dozens of emails, mostly generated by AI, I'm sure, at this point.

1:42:11Hackers aren't that ambitious. Okay. All right. On we go. So what is weight space programming? We all know that this podcast is not directly about AI. Twit has Intelligent Machines podcast for that. But AI is obviously rocking the boat of everything we do on this podcast, regarding cybersecurity. And so for that reason, and just for the sake of my own edification, I previously invested a bunch of time working to understand AI as much as I could and sharing what I've learned about it along the way. So when something truly fundamental and significant happens in AI, since we now have a foundation for putting such events in context and understanding them, I'm going to take a moment to keep everyone current.

1:43:04Such a thing recently came to light with the publication of research out of MIT's AI lab, which has made a rather amazing AI-enabled breakthrough in understanding and deliberately modifying the hundreds of millions of weights within full-size neural networks. As we've covered before, pre-training is the process of training a model to predict next tokens by using gradient descent to tweak the values throughout a huge network to improve its prediction. That process ingests a massive volume of language and then after training, well, I'm sorry, ingests a massive volume of language. And then after training, which requires millions of dollars of compute, the result is a large model that models the language contained within the corpus it was trained on.

1:44:10But that model does not yet have any behavior. So various forms of post-training are then used to imbue this model with behavior. Before that, it is behavior-free. It's just a statistical model. But this behavior turns it into a conversational chatbot, which produces response that humans and sometimes other already trained models judge to be best as part of this whole feedback mechanism to keep training this. Okay, so the most important takeaway here is that at no point in any of this is there any attempt to understand the values of the many hundreds of billions of individual high-precision floating point weights inside the model.

1:45:10We just stand back and the magic somehow happens. Even when ablation is used to remove behavior that has been imprinted onto a model through what's known as low-rank adaptation, low-rank because not that many weights are changed, those removing unwanted behavior, such as a model's refusal to answer some questions, they do that by observing the activations within a model while the model is refusing and then imposing a nullifying change to keep those activations from occurring. But again, no attempt even there to actually understand the why of the weights. The model is simply forced to reveal them, the important ones, to that refusal, at which point they can be nullified.

1:46:12What MIT has successfully accomplished is an AI-enabled breakthrough in large language model weight understanding. They are using one AI model to understand and deliberately alter the weights in another. In their research paper, the MIT researchers demonstrate that LLMs are surprisingly competent at understanding and deliberately editing the weights of another network. To me, this seems likely to keep the AI doomers up at night. I'm going to share just the TLDR at the top of the paper, which they posted last week, just to give everyone a sense for this. So the four MIT AI researchers wrote, we find that frontier LLMs are surprisingly good at natively reading and writing neural network weights, much like code.

1:47:22sometimes with tools such as a python kernel and sometimes from the raw weights alone. These results show sparks of weight space programming, which may extend beyond what we'd expect from humans or conventional algorithms. This might be a big deal for both optimization and interpretability. Just by looking at the raw weights, GPT-6 Astra can tell what a network was trained to do. And without any environment feedback, it has no trouble writing from scratch small game playing networks that outplay networks trained with PPO, which stands for proximal policy optimization, which is a technology that I think it was OpenAI came up with for this sort of for like game playing optimization.

1:48:24So no problem writing from scratch small game-playing networks that outplay networks trained, essentially classically trained networks, for 100 million steps of gameplay training. They said,

1:49:110.016 % of the weights. With the same sampling settings, this weight correction raises Quinn's performance on the Math 500 benchmark from 7.9 % to 48.1 % and generalizes out of the box to instruction following, raising IF eval, which is the instruction following eval benchmark, by a factor of three. All of this is possible without Astra to write training loops. It uses its reasoning to figure out what each part of a neural network computes, or should compute, and changes the weights accordingly, much as a programmer reads and edits code. We call this concept weight space programming, treating the weights as a program to be read and written.

1:50:15Throughout the history, they conclude, of deep learning, we encoded our simplistic inductive biases into tiny symbolic optimizers like atom or muon. This is changing. Our neural network weights will soon be natively optimized by LLMs that deliberately reason about their weights. And that's the beginning of a long blog. I've got the link to it in the show notes for anyone who's interested. And if a chill just went down your spine, that's because you've understood the essence of what this means. LLM technology is far beyond human understanding at the level of its weights. Hundreds of billions of parameters.

1:51:11We can't even truly conceive of that quantity, let alone the meaning of the various values of that many weights. You know, yeah, we managed to create it and then we trained it through a process that's sometimes referred to as growing an AI. But by their own frequent admissions, even those who do this do not fully understand how their own creations work. We use what amounts to brute force training to bootstrap an LLM into something and no one is sure exactly what. But out of that, we then managed, we have managed, MIT has managed to cobble together a reasoning model. And it turns out to be able to reason about its own model weights, which, as MIT researchers have shown, can be used to improve the weights carried by existing models and to even create outperforming existing outperforming models from scratch.

1:52:21So anyway, there is a lot more in their paper and in the news reports. So as I mentioned before, there's a little bit of controversy about this, which is interesting. But that always happens when you shake the foundation that's already been established. And this is foundation shaking. Many of the people who have seen this believe that it will change everything. And to me, it makes sense that it will.

1:52:51A brief note on the Spinrite front, what we're calling Spinrite on UEFI, which will become Spinrite 6.2, that project is proceeding with gratifying speed. When I wrote this yesterday, I asked Claude, and it's difficult for me to, you know, not say who is serving as my project manager. Claude is serving as my project manager for this work and helping me to keep track of everything. I asked for a count of machines that have been reported to successfully boot into Spinrite from UEFI firmware versus those that we are still working on. And this is now a day out of date and things have been moving so quickly that the numbers are no longer any good.

1:53:46But essentially, 44 in total, completely heterogeneous collection of machines are successfully booting. And among our testers, four we still have are tripping us up. And there's something about the late model AMD Ryzen processors, which we've not yet nailed. But this is really moving quickly and the testers are having fun. And essentially, there's no showstoppers. I expect that when I check in next week, I wouldn't be at all surprised if we're running on everything everybody has. So we're just about there. And I did want to share one piece of feedback because there's an interesting inversion that surfaces from this question that although the answer is kind of obvious, the way of thinking about it, I think, is interesting.

1:54:45A listener of ours, Eric Michael, he said, Greetings, Steve. I listened to episode 1103 in your commentary on encryption fast mode, where every third block of a file is encrypted. How is this fundamentally different from modern ransomware, which frequently encrypts only portions of files because that's sufficient to render them unusable? We do not see tools capable of keyless recovery of ransomware encrypted files, despite substantial portions of those files remaining unencrypted. He said, I agree that fast mode does not provide the same cryptographic protection as properly implemented full file encryption.

1:55:35However, if the practical objective is to prevent access to usable files without the encryption key, does encrypting every byte actually provide a meaningful advantage over encrypting just enough to make the files unrecoverable? Ransomware operators have effectively demonstrated that partial encryption can achieve that objective. It seems to me that the same principle applies to fast mode, even if it falls short of full encryption from a cryptographic standpoint. I look forward to your thoughts, longtime listener and Spinrite owner, Eric. So, okay. I think that there's an interesting lesson that a thoughtful answer to Eric's question can teach.

1:56:35The reason partial encryption in the ransomware scenario differs from partial encryption in the personal data storage instance is the differing needs of their users. The need in the ransomware scenario is for the data's rightful owner to recover all of the lost data, whereas the need of the owner of data in the personal storage instance is to prevent any of their data from leaking. And those are very different requirements. We don't know how large the individual blocks of data might be. where this thing, this fast mode, all we know is that it's every third block it encrypts. Any key length of AES cipher uses a 128-bit or 16 8-bit characters block.

1:57:38But the lazy encryption or fast encryption might have been by storage media sector blocks, So, you know, 512 byte sectors. But regardless, looking at someone's archive, for example, of email, where only one third of the total was gibberish, that would still reveal a great deal of potentially compromising information and would rightly horrify anyone who was assuming that they were being protected by, you know, military grade whole device. encryption when in fact that was not there. So I do think that there is a difference between a ransomware bad guy encrypting a third of a victim's data and someone saying, well, you could run this encryption fast.

1:58:34And if it's only a third being encrypted, there's still a lot of information leakage. Okay. So Leo, let's take our final break. And then we're going to dig into another, we already talked about AI and this potentially groundbreaking innovation from MIT. Here's another one. The idea that we could actually finally prove to a mathematical certainty that code is correct and bug-free and does what we want. Wow. What a show this has been. We'll get to that in just a bit. But first, let me make a plea to all of our fabulous listeners, those of you listening right now, to join Club Twit. We're coming to the holiday season and this is a great time to do it.

1:59:30Club Twit makes a huge difference in our bottom line. Without it, we would have to cancel even more shows. We've had to cancel a few. I hate to say it, but we just don't have additional money that we can pull out of thin air. Without your help, you know, this operation would be struggling quite a bit, but you pay for about 40 % of our overall operating expenses. That's great. And we'd like to have more of you. We'd be less dependent on advertising, which is kind of a seasonal come and go business, and be able to do more of the kind of programming you want. It's a way of voting in a way for what you like.

2:00:06if you like what you hear and you want to hear more twit.tv slash club twit you get ad-free versions of all the shows you get chapter markers and all those shows you get uh transcripts you can read along while you listen uh we've got of course the great club twit discord and all the special programming we put up put on in the club to a discord we uh we really love what we're doing and we want to do more of it if you love what we're doing the best way to support it go to twit.tv slash club twit as we get close to the holiday season. This is a good time. We've got a 10 % discount for the annual membership.

2:00:40Great for gift giving or make yourself the recipient. You could do that. Some great stuff coming up this Friday. Photo time with Chris Marquardt at 1 p.m. Pacific off by one right after that. 2 p.m. Pacific with Jeff Atwood, Micah's Crafting Corner next Wednesday. A week from Wednesday, our user group is back December 4th. The AI user group, the Micah Media Club. Still working on what that will be all about January 8th. Just some of the things coming up. Twit.tv slash club twit. Help us help you by joining the club. We appreciate everything you do to keep us on the air. Now we continue on, ladies and gentlemen.

2:01:22Join the club. One of us with Security Now and, wow, the case of the provable code. So, yes. Last Tuesday, a week ago, November 10th, Google DeepMind published a research post titled Introducing Alpha Proof Code, a breakthrough in verifying complex software at depth. Okay, so a little bit of background. And this is, I warn you, going to kind of be a conceptual thinking piece here. For anyone who has spent their career around software, what Google is claiming reads almost like science fiction. AI agents guided by a small team of human experts. produced a true mathematical proof that the actual compiled machine code, because that's what it didn't use source, it used the machine code of signals, core cryptic, and core cryptic, and core, secrecy.

2:02:46These were implemented in two parts of LIBE signal, the PQXDH handshake, which we talked about previously when signal added quantum to their existing Diffie-Hellman, and the triple ratchet. And again, as I said, this was all done. This was validated at the compiled machine code level. I'm not going to get into the details of those four security principles, since even though it's terrific that what was believed to be true and correct of them has now been proven to a mathematical certainty, the real story here is that extremely complex cryptographic code could actually be proven to be correct, which is huge.

2:03:40And it's, well, we know about bugs, right? It's unimaginably important. The other nuance worth mentioning is that it was not Signal's various protocols, nor even its source code, as I said, that was analyzed. the proving system worked from the machine code that was running on the phone that Google used. This allowed its proof. Oh, and this is a static analysis. It didn't run the code and watch it. It dumped out the program and then analyzed it. So this allowed the proof that they were able to generate to encompass not only the protocol and not only the source code that was used to implement the protocol, but also the compiler that was used to turn the source code into machine language, which the machine would then execute.

2:04:37So it's as low as you can go, and it cuts out all of the intervening opportunities for there to be an error. And it was not simple to do this. The proofs that were developed total about six and a half million lines using the lean proof language. A five-person team accomplished this in four weeks, and Google's point of comparison is that verification at this depth has historically taken teams of experts years when it could be done at all. Okay, so this brings us to what I want to spend the time on today, which, why has this been so hard for so long? I guess the idea is kind of obvious, right? That is the idea of doing this, the dream of proving, proving, meaning mathematically with certainty, proving the correctness of a program.

2:05:42Well, it's as old as the world of programming itself. Yet for 60 years, that dream has been stubbornly, you know, inaccessible. It just in the case of real world software. And when you stop to think about it, the idea is not even on anyone's radar. We don't talk about it, right? I mean, it's not something, it's like, oh, we got to go find the bug. There's, it's not even like, there's no idea that it's possible to prove something is bug free. It's just so out of reach that no one even teases the idea. And, you know, I don't know if I can even explain how much I dislike the glibly rattled off statement, which we hear, you know, and I mentioned before, all software has bugs.

2:06:31You know, I guess the reason it irks me so much is that it's used as a justification and an excuse for giving up on making code as perfectly as it can be. You know, the sentiment is, well, it's good enough and we'll fix it when problems are discovered in the field, which is the experience we all have. Okay, so what's the problem with formal code verification that has made it so difficult that no one even considers it a possibility? I mean, except way out. Everything we do to gain confidence in code, when you think about it, unit tests, code reviews, static analysis, fuzzing, penetration testing, bug bounties, all of that is actually a form of sampling compared to rigorous proving, right?

2:07:22It's a form of sampling. We try some inputs and we see what happens. A famous computer scientist, Edgar Dijkstra, expressed the limitation perfectly back in 1970 when he said, program testing can be used to show the presence of bugs, but never to show their absence. What do you think about it? I mean, that's the case. Testing can be used to show the presence of bug. That's what we do with fuzzing and with various types of testing. But how do you prove a negative? How do you show their absence? And the problem is that things get complicated very quickly as we begin to scale. You know, a function that takes, for example, just as a pure abstract, two 64-bit values, therefore has 2 to the 128 possible inputs.

2:08:24And if we just brute force this, if we tried to test a billion of those possibilities per second, we would still need on the order of 10 to the 22 years to try them all, which is about a trillion times the age of the universe. So that's not going to work. Of course, real programs take kilobytes or megabytes of input and carry huge internal state, so vastly more complex than the idea of some function that takes two 64-bit values, which makes that kind of exhaustive testing, you know, not merely impractical. It's not even a meaningful concept. You can't even get there. As we have often observed on the podcast, this is the big asymmetry of security.

2:09:20The defender has to get every case right. The attacker needs to only find one input that nobody tried. Every buffer overflow, every use after free, every parsing bug, which we've covered on this podcast for 20 years is at its root an input that testing never sampled. That's what that means when you find something that allows you to bypass the authentication on your Synology NAS. We never hit that in testing. Otherwise, they would have removed it. So, it was sampled. a proof so is different in kind not just in degree a proof doesn't try various inputs it reasons about all of them at once you know sort of the way you think of it the pythagorean theorem holds for every right triangle without anyone measuring even one we know it's true We can see it.

2:10:26We can see how it's constructed. If you could prove that a program does what it is supposed to do, nothing more, nothing less, for every possible input, then entire classes of vulnerabilities would simply cease to exist for that code. They wouldn't become rarer or more difficult to find. That's the hill we're climbing right now with AI vulnerability discovery, right? You know, it's like, oh, you know, and we just saw last week we got a stronger GPT and it found more problems that the previous one hadn't. So but that's this is different. That's why this is so fundamentally, I mean, in computer science, it is earth shattering.

2:11:18It changes everything. It is the holy grail. It has always been the Holy Grail, and it's not been available. So why not? Why hasn't anyone grabbed it? The first obstacle predates computers even. Back in 1936, the brilliant mathematician Alan Turing proved that no algorithm can examine an arbitrary program and decide whether it will ever halt. It's weird to say that, but it's true. No algorithm can examine an arbitrary program and decide whether it will ever halt. Then in 53, Henry Rice generalized that into what's now called Rice's theorem. Any non-trivial question about what a program does, will it ever overflow its buffer?

2:12:14Will it ever leak a key? Does it compute the right answer? Is undecidable in general. So the fully automatic dream, feed in a program, press a button, get back correct or buggy, is mathematically impossible. Not hard. It's impossible. So say these early computer scientists. But there is a loophole in that argument, and it's the key to this holy grail. Rice's theorem states that no machine can find the answer for every program. It says nothing against a machine's checking a proof that someone else found for one specific program. Finding a proof can require insight, you know, and some creativity and maybe lots of searching.

2:13:17But checking a proof that exists, that's mechanical. Does each step follow from the previous ones, given the rules of the proof system? So a proof checker can be a small, simple, carefully audited program that can be known to be true. In fact, I suppose you could have it prove itself. So proof checking is easy, but generating the proof to be checked, that's the hard part. And this has been on the minds of mathematicians. In 1967, Robert Floyd wrote Assigning Meanings to Programs, and a couple years later, Tony Hoare published An Axiomatic Basis for Computer Programming. Those works, those two pieces, provide the core idea which is still used today.

2:14:17Attach logical assertions to a program. This must be true before this statement runs, and this will be true after. So then you prove each step preserves those assertions. In Hoare's notation, it's curly braces around a P, then a C, and then curly braces around a Q. And that became the foundation of the field. And at the time, there was real optimism that programmers would soon, this was back in 67, soon routinely prove their code correct by the way mathematicians prove theorems. Not so easy. Ten years later, in 1979, a paper appeared in the Communications of the ACM titled social processes and proofs of theorems and programs.

2:15:12The argument made by its authors was that mathematicians don't believe a theorem because a proof exists. They say, well, okay, maybe. They believe it. They believe the theorem because a proof exists because and only when other mathematicians have read the proof, argued about it, simplified it, and used it in order, in other words, sort of proving the proof. Program correctness proofs, they argued, are enormous at, well, six and a half million lines of lean code, enormous, tedious, and here's another problem specific to just one piece of code. Who's going to read that? Nobody. And an unread proof, well, it's not able to convince anyone.

2:16:09And that's what we've been seeing with these breakthroughs in math, right? It's like, it's, you've got to read through this proof and, you know, try to understand it and then say, okay, that, that works. So, and as we know, because that was then and it hasn't happened. History proved them right. In 1996, Hoare himself gave a talk with a title titled, How Did Software Get So Reliable Without Proof? And the answer to that question is the industry had found that testing, review, and good engineering was good enough most of the time and at a tiny fraction of the cost. In other words, sampling, just sampling some parts of the program's operation.

2:16:57It's like, well, okay, it seems to be working. And that was not expensive to do. Proving, so just like in a different dimension of expense. Okay. So this brings us to the question, what is it exactly that we want our proof to prove? In an absolute sense, you know, that a program has no bugs or that it always does what we want, which brings us to the issue of needing a functional spec. In 1977, Donald Knuth ended a note to a colleague, and this was famous. I'd seen this a lot. Maybe our listeners will be familiar with it. It's become one of computing's greatest one-liners. He said, beware of bugs in the code above.

2:17:45I have only proved it correct, not tried it. And so, you know, that was also Canoes' famous wit. But he was making a serious point. A proof never establishes that code is correct in some absolute sense. It establishes that the code matches a specification, a precise, formal statement of what the code is supposed to do. And writing that specification is itself a programming problem with its own bugs. Prove the code meets the wrong spec or an incomplete spec, and you have a terrific proof proving the wrong thing. For something like a sorting routine, The spec is simple. The output is a permutation of the input in order.

2:18:41But what is the formal specification of, you know, a secure messaging protocol look like? We need to express mathematically exactly what an attacker can and cannot do, what it means for a key to stay secret, what happens after a device is compromised and later recovers. Cryptographers spend their time working to get those definitions right. So, with Google's DeepMind, those five guys spent their time carefully reviewing every specification for signal. They were able to use AI to prove things, but it was they who needed to first decide what it was that it needed to prove. Now, we've often talked about the shockingly high cost of creating ultra high quality software.

2:19:43Software, you know, that runs mission-critical systems like the U.S. Space Shuttle program, where bugs must absolutely not be allowed to cause critical harm. Creating perfect software is so surprisingly expensive that it is almost never done. There are a couple examples. SEL4. Back in 2009, a team at Australia's NICTA, their National Information and Communications Technology Research Center, they completed the first full functional correctness proof of a general purpose operating system microkernel. The kernel was described by around 8 ,700 lines of C plus, about 600 lines of assembly. The proof in something known as the Isabel HOL theorem prover ran about 200 ,000 lines.

2:20:56So again, the kernel that they wanted to prove the correctness of, 8 ,700 lines of C +, about 600 lines of assembly, 200 ,000 lines in this theorem prover dialect. It took 20 person years of effort by world-class experts to produce that specification. And because the proof is tied to the code, every change made to the microkernel means revisiting the impacted portions of the proof. But in return for this gargantuan effort, SEL4 is a true triumph. It's used in defense and aerospace systems today. Its Wikipedia entry says,

2:22:22prove the system's confidentiality, integrity, and availability, among other properties. The initial paper outlining SEL-4's verification was inducted into the 2019 ACM SigOps Hall of Fame. The kernel is extremely minimal, handling only memory management, process isolation, and scheduling. Its inter-process communication acts as a protected procedure call for cross-domain invocation rather than general-purpose message passing. SEL was developed as a from-scratch microkernel design influenced by the L4 microkernel family with an explicit goal of enabling comprehensive formal verification while maintaining high performance.

2:23:15In 2009, the SEL4 project reported a machine-checked proof of functional correctness spanning from formal specification to C implementation. So, one thing to note is that that work, which was heralded as a monumental breakthrough in code verification, only verified that the C source code correctly expressed an implementation of the specification. That's what was being verified. That's what's different about what Google has accomplished. their effort went another layer deeper, verifying that the compiled code, regardless of its implementation language, you know, the actual machine language that the processor silicon would be running perfectly and correctly implemented the design's specifications.

2:24:16So is going down to the bare metal actually important? Well, maybe. An effort known as CompCert is a formally verified optimizing compiler for a large subset of the C99 standard of the C programming language. Researchers at France's INRIA, their National Institute for Research in Digital Science and Technology, invested years building and proving a C compiler using the ROC, R-O-C-Q, prover. ROC describes itself as a general-purpose, industrial-strength, interactive theorem prover for mechanized reasoning in mathematics, computer science, and more. Using this ROC and their CompCert compiler implementation, the French researchers were able to demonstrate that the machine code it, the compiler, produces behaves exactly as the source program says.

2:25:27Then, the value of that work was demonstrated a couple of years later in 2011 when researchers at the University of Utah built a tool called CSmith, which generated random C programs to hunt for compiler bugs. They found hundreds of bugs in GCC and LLVM, obviously two super useful and important compilers. But in CompCert's formerly verified compiler, they found none. So there's still another problem with compilers that I remember we've discussed in the past, which is that allowing a compiler to optimize code might not change the program's behavior, but it might break important security guarantees.

2:26:19Ken Thompson made the famous version of this point in his 1984 Turing Award lecture, which he titled Reflections on Trusting Trust. He showed how a compiler could be made to insert a backdoor, made to insert a backdoor that appears nowhere in any source code that could be inspected, and cryptographers have understood and fought the problem for decades. Following best security implementation practices, a careful programmer, say, for example, deliberately zero fills a buffer, which was holding a sensitive cryptographic key. You know, and that's done the moment it's no longer needed. The optimizer notices that the buffer is never read again and quietly deletes it because it's obviously not needed.

2:27:16It can't be needed. Then code which has been painstakingly written to run in constant time so that it will not leak secrets through side channel timing attacks. Well, it becomes improved, I've got an air quotes, by the compiler into code that now has data dependent branches. An examination of the source code, it looks perfect. But the binary that the compiler produces is not, because it optimized the program and made it insecure. We've seen that such side-channel timing attacks are far from theoretical. All of those processor optimizations that resulted in the specter and meltdown vulnerabilities came back to bite us, as we know.

2:28:03So this is why DeepMind's choice of optimization target, which is to say the actual post compiler compiled machine code matters so much. Signal's LibSignal crypto library is written in Rust. That's a good start. And as DeepMind's authors note, verifying at the Rust source level means trusting Rust-C, LLVM, and a formal model of Rust's semantics. So they chose to instead prove the properties of the compiled ARCH64 machine code. Just for the record, there have been other successful efforts at formal code verification when it really mattered. That infamous, remember the 1994 Pentium floating point, you know, the FDIV bug, that was caused by a handful of missing entries in a lookup table.

2:29:06The cost to Intel in 1994 dollars was reportedly around 477 million dollars. After that, Intel and other chip makers invested heavily in formally verifying their floating point and other logic. So it's doable, it's just hard as heck and expensive. Hardware, however, is well-suited to formal verification. Designs are finite, specifications are precise, and it's important because a bug cannot be patched once the silicon ships. So getting the design right before fabrication will pay off even against a heavy headwind of expense. Formal verification has been used against critically important cryptographic primitives.

2:29:57Various labs, including Microsoft's, France, INRIA, again, and MIT, have produced formally verified implementations of the low-level math underlying TLS, thank goodness, and other protocols. Verified code from those efforts lives in Firefox, in Google's Boring SSL, and therefore in Chrome, and in the Linux kernel. They're all using formally verified instances of code. Amazon Web Services built an automated reasoning group that uses solvers to check things like IAM access policies and pieces of its own reduced size TLS implementation. We talked about AWS's TLS, which they rolled for themselves, you know, pulling forward none of the legacy crap that OpenSSL has and how much smaller it was.

2:30:52And before DeepMind came along, the signal protocol was already one of the most formally analyzed protocols in existence. Academic teams have analyzed its design with symbolic and computational tools like ProVerif and CryptoVerif since at least 2016. And signal's newer post-quantum designs ship with those formal models. However, importantly, that previous work did pay off and revealed shortcomings in the earlier implementations, which then led to their resolution. Okay, so the thing that is different about Google's deep mind accomplishment is its scale. And as we've seen, scale matters significantly because the complexity of the work escalates quickly at scale.

2:31:48Earlier work has always either been against a deliberately small kernel of code or an abstract model of a design. They were deliberately kept small because they could not be handled if they were any larger. As DeepMind notes, most protocol verification stops at the design level. The analysis proves the protocol on paper is sound, and then one or more people would write the real code, hoping to make it match the spec. Right. Until now, nobody had been able to similarly prove the operational security of any widely deployed protocol's actual running implementation, because the cost to do so would just be stratospheric.

2:32:35This brings us back to what's been known about this problem, right? Right. Generating a proof is the hard part, but verifying what is claimed to be a proof is comparatively easy. Historically, generating the proof required expensive, slow and painstaking human experts. DeepMind's work has shown that AI can take over the task of proof generation to make practical, much more sizable successes in formal code generation. Now, AI making a proof? Anyone who's worked for long with a large language model will have experienced their well-known characteristics of being fluent and confident, but sometimes wrong.

2:33:28Having one's AI produce a lovely legal brief, complete with great-looking but hallucinated prior case law references, has landed more than one attorney in hot water. But an interesting characteristic of the use of LLMs to aid formal code verification, and this is really interesting, is that for theorem proving, being wrong doesn't really matter. The reason is that the AI is generating the proof, and that does not need to be perfect. Every proof it produces is then handed to a proof checker, in this case, the kernel of Lean, which is the proof assistant which Microsoft Research created back in 2013.

2:34:24Lean will only accept a proof where every single step in the proof is valid. Consequently, no hallucinated proof is able to slip through. The proof will simply fail to check, and the AI will try again. The model, in this case, can be wrong a thousand times. It only has to be right once, and the checker tells us when that happens. D-Mind has been working on this for a while. In July of 24, its alpha-proof system, working in lean, reached a silver medal equivalent score at the International Mathematical Olympiad, solving problems that stump most of the world's best young mathematicians. Alpha Proof Code takes that same proving engine from competition math to production software.

2:35:20And this brings us back around to that 1979 paper published in the ACM, Social Processes and Proofs of Theorems and Programs. Those three authors argued that mathematicians don't believe a theorem just because a proof of it exists. They believe it because other mathematicians have read the proof, argued about it, simplified it, and used it. They noted that any program correctness proofs would be enormous, tedious, and specific to only that one piece of code, thus essentially rendering it, you know, the whole concept impractical. I would have titled today's podcast Practical Bespoke Proof Verification, but no one would have known what to make of that.

2:36:09However, that's what just happened. Practical Bespoke Proof Verification. All right. So combining a tireless AI-powered proof generator with a trustworthy and equally tireless proof correctness verifier, and we have another brand new AI-enabled capability that has never been remotely practical before. So that's the story of what Google has invented with this, you know, this alpha code proof project. It's not clear where we're going to go from here. What it does open the door for is AI now, as long as you have a specification, that's you still need to specify what you want to have done. Obviously, writing an exhaustive spec for Excel, for example, I can't even imagine something like that.

2:37:08It was difficult enough to create a specification for what signals cryptographic algorithms do. But they did because it was important. What this allows us to do is take other large, important pieces of code and have an AI and create a spec for it. And oftentimes the specs are not complex, whereas the code is. create a spec for it, and then something like alpha proof code will be able to attempt to prove it by creating a proof which the checker will be able to say, yes, that is an absolutely correct proof or not. And if not, the AI will try again until it gets it right. So a very new piece of computer science technology that wouldn't have been possible without AI, because it turns out AI is able to turn specs into, hopefully, into correct proofs.

2:38:12And we know that the AI is going to get better over time. So this is really a cool approach to tackling the problem. Yeah, very interesting. Very interesting. Do you think ultimately it'll mean better code? it's going to mean better code where it matters yeah so so you know all we've done all we've been able to do so far is inspection and fuzzing and trying it and seeing if it works you know various types of sampling where in instances where something has to be absolutely correct where you can't tolerate a bug, then as long as you can specify how you want it to behave, we now have the way of verifying that, yes, that code absolutely behaves to the spec.

2:39:02Right. Cutting out all uncertainty of human programmer or compiler error. Yeah. Well, very interesting stuff. A lot of speculative stuff in here and a lot of hardcore, real-world, crashing kind of stuff. This is a good mix for today's security now. Steve Gibson is at GRC.com, the Gibson Research Corporation. That's where you'll find Spinrite, the world's best mass storage maintenance recovery and performance-enhancing utility. You should have a copy of that in your library, along with his DNS Benchmark Pro, which helps you find the best DNS server for your particular location. both at grc.com if you want to send steve comments suggestions or pictures of the week first you have to whitelist your email by going to grc.com slash email while you're there you know submit your email address below it there are uh you can sign up for two different mailing lists one is the show notes that come with every show he spends a lot of time 20 or so pages every week of great stuff you can get that emailed to you on sunday or monday before the show you can also sign up for his new product announcement now that you're using ai steve i expect a new product every few days we'll see at least we'll know that'll be provably correct better than every few years yeah every decade uh grc.com you can also when you're there get a copy of the show he has unique versions of the show when it comes to audio 16 kilobit audio for the bandwidth impaired 64 4KBit for those who know what they want, I guess.

2:40:44And the show notes are there for download. And transcripts written by an actual human, Lane Ferris. Hi, Elaine. Those come a few days after the show concludes. We have on our website, we have a, I think it's 128-bit MP3 version of the audio. But we also have video, which no one else has, twit.tv.sn. There's video on YouTube. There's a dedicated channel. Great for sharing clips. And of course, the best thing to do is subscribe on your favorite podcast player and you can get it automatically as soon as we're done. You can even watch it if you want the freshest version as we do it live. We record the show immediately after Mac Break Weekly between 1.30 and 2 p.m.

2:41:24Pacific. That's 4.30 Eastern. That's 21.30 UTC. You'll find us on all the streaming channels. Of course, club members get to watch behind the velvet rope in the Discord, but there's also YouTube, Twitch, X, Facebook, LinkedIn, and Kik. So pick your poison and watch live. We love having you on Tuesday because I watch the chat. And so it's nice to see the feedback as we go through the show. Well, that's about it, Steve, for this Tuesday. Next week, another exciting edition. Who knows what will happen between now and then, Leo. It's Thanksgiving edition. It's going to be wild, I tell you. It feels like there's more.

2:42:04the other shoe is going to drop or something. I don't know. We're living in interesting times. See you then, Steve. Wouldn't have it any other way. Bye.

2:42:17Security now.

2:42:24It was many nights before Christmas and one woman knew she needed a Christmas tree so to Wayfair she flew. Finding just the right one and for years it will last. Wayfair shipping was free and remarkably fast. What's this? Garland? A wreath? Essentials to host? Good thing she celebrates way longer than most. As her cart filled with magic and decor from the heart, Merry Christmas to all. Get your Wayfair head start. For holiday decor you can count on year after year, Wayfair delivers.

From the publisher

What if AI could finally prove your code is bug-free? A new advance out of MIT and Google DeepMind is redefining what's possible in software verification—and it could mean a future with provably correct code, not just wishful thinking.

  • Last week's Microsoft 365 and FAA nationwide ground stops.
  • Riding the Patch Tuesday rollercoaster.
  • Speaking of Patch Tuesday: the return of Nightmare Eclipse.
  • Synology patches a critical 9.8 remote authentication flaw.
  • The unexpected revocation of GRC's code signing certificate.
  • "Weight-space programming" and its implication for AI.
  • SpinRite for UEFI progress report.
  • Why only encrypting every third block is not okay.
  • The holy grail of Formal Code Verification

Show Notes - https://www.grc.com/sn/SN-1105-Notes.pdf

Hosts: Steve Gibson and Leo Laporte

Download or subscribe to Security Now at https://twit.tv/shows/security-now.

You can submit a question to Security Now at the GRC Feedback Page.

For 16kbps versions, transcripts, and notes (including fixes), visit Steve's site: grc.com, also the home of the best disk maintenance and recovery utility ever written Spinrite 6.

Join Club TWiT for Ad-Free Podcasts!
Support what you love and get ad-free audio and video feeds, a members-only Discord, and exclusive content. Join today: https://twit.tv/clubtwit

Sponsors:

More from Security Now (Audio)

All episodes