Mythos & AI Bad Behavior 4/9/26

15:13Michael Froman:After all, I think that was the bet by a lot of people. It was even a Tuesday, wasn't it, when it happened? So everybody, we know what Tuesday is.

15:21Dave Kasten:Taco Tuesdays. I actually made tacos on Tuesday.

15:25Michael Froman:You did?

15:25Dave Kasten:I did.

15:26Michael Froman:You know, any day is a good day for...

15:28Dave Kasten:Tacos. A federal appeals court denying Anthropics' request to temporarily block the Pentagon from blacklisting it as a lawsuit challenging that would ban it plays out. The case is one of two, and it's a little complicated, one of two related to Defense Secretary Pete Heggs has moved to designate Anthropics a supply chain risk after the company sought limits on how the technology was used. Late last month, you might remember in another case, a federal judge ruled for Anthropics, saying it appeared the Pentagon retaliated against the company with the split decision. Split decisions now by the two different courts.

16:04Dave Kasten:Anthropic will be excluded from Defense Department contracts for now, but is able to continue working with other government agencies while the litigation continues. Katie Hahn, who I remember we had a conversation with about this, effectively took to Twitter last night, or X, and said that the three judges that were picked to choose this case, they said the administration couldn't have gotten a better group of people for their side if they had selected them themselves. There's three out of the 11. And so the question that she was sort of raising was, if you could try to go back to the court, which they could do, and do what they call en banc and try to get all 11 judges to try to overrule it that they, based on their backgrounds and things, that they would.

16:54Dave Kasten:So it's going to be very interesting to see what happens ultimately. I didn't realize that part of it, but the government's case, or the judge's case in siding with the government, was that this is the fortunes of a private company, a single private company, versus the ability of the Defense Department to decide what it thinks is important during times of war, which was kind of an interesting argument. I hadn't thought about it from that perspective. But they basically said they didn't care about the fate of a single private company. Right. What was interesting, though, to me, and this is the thing that I think where like a Katie Hahn would come out if you were looking at this just in terms of what's actually at stake, which I didn't understand the way the judge described it.

17:37Dave Kasten:You don't have to determine that it's a supply chain risk to not use the service. So the government could say we are not using Anthropik. That is their choice. It's the designation of them as a supply chain risk, which then has all of these downstream effects to the company. And so when you look at why the other judges on the West Coast said that it was retaliatory, it was retaliatory because it effectively, by describing it as a supply chain risk, then hurts it in these other ways. It's not a final ruling, right? This is just to say that the injunction, that they're not going to give relief while they're deciding the rest of it.

18:15Dave Kasten:Right. And so now we'll go to court. By the way, it'll go to court in California. It'll go to court in D.C. as well. And then we'll see. And I don't know if these justices are the same justices that will actually determine that outcome.

18:28Michael Froman:I can ask you off camera. Okay. Was it ground beef tacos, chicken tacos? Ground beef. Was it ground beef? It was.

18:35Dave Kasten:Sometimes I make turkeys. Hard show? Soft.

18:37Michael Froman:Soft show. You put salsa? No. No salsa? Sour cream. Visualizing. I need info. On tacos.

18:46Dave Kasten:in particular. Cheese will be next. Next on Squawk Pod, the fog of the Iran war with the Council on Foreign Relations President Michael Froman. Who is beholden to the ceasefire and who are the stakeholders in a solution?

19:03Michael Froman:We would be in a stronger position if we had allies with us. Most of the oil doesn't come to the United States, goes to Asia, and to a certain degree to Europe, but we are not unaffected.

19:13Dave Kasten:We'll be right back.

19:18Dave Kasten:Is Bitcoin complicated or do people just think it is? For a long time, it felt like something only tech experts or hardcore investors really understood. Between all the headlines, new terminology, and everything people say about it, it can seem harder than it needs to be. But what if getting started with Bitcoin didn't have to feel overwhelming? If you've been curious about Bitcoin but haven't made the jump yet, Cash App makes it easy. You can set up automatic purchases with zero fees or buy larger amounts, also with zero fees. Start small or go bigger. It's designed to be simple either way. For a limited time, new customers get$10 added to their balance.

19:56Dave Kasten:Just use code CASHAPP10 when you sign up and, don't forget this part, send at least$5 to a friend in the first two weeks. Terms apply. Cash App is a financial services platform, not a bank. Banking services provided by Cash App's bank partners. Bitcoin services provided by Block, Inc. Brandt. For additional information, see the Bitcoin disclosures at cash.app slash legal slash podcast. Never bet against American grit or American energy. Through innovation, Venture Global is not only building some of the largest energy facilities in the world right here in the United States, but delivering American energy at a fraction of the cost and a fraction of the time.

20:36Dave Kasten:So while others are busy talking, we're busy building. That's Venture Global. That's unstoppable energy.

21:06Dave Kasten:of everything. Help kids save, budget, and build financial independence all with Chase. That's good for kids and good for parents. Visit your local Chase branch and get started. Accounts subject to approval. JPMorgan Chase Bank and a member FDIC. Welcome back to Squawk Pod from CNBC with Joe Kernan, Becky Quick, and Andrew Ross Sorkin. Here's Andrew. President Trump announced a ceasefire in the Iran war about 36 hours ago, but things have only gotten more complicated since then. For more, we want to bring in Council on Foreign Relations President Michael Froman. He served as U.S. Trade Representative in the Obama administration.

21:43Dave Kasten:Good morning to you. Morning, Andrew. You know, I think we, it's not like I want to say we said this already, but, you know, everybody seemed to, you know, think that things were so great yesterday. I was not so sure why that was the case. And the real question now is, is it one step forward, two steps back, three steps back, no steps back? What are we?

22:03Michael Froman:Well, first of all, I don't think it's unusual when these ceasefire agreements are announced that there is some bumpiness in the road in how they're implemented. As orders are given down the line, whether commanders in the field are being effectively sort of kept under the control of a central authority, etc. So it's not unusual for there to be at least some period of implementation noise in the system. I think the problem here, Andrew, is that there doesn't seem to be total agreement as to what's included in the ceasefire or not. Israel's actions in Lebanon, are those included? Did the U.S. accept, did the president accept as a basis for negotiation, a 10-point plan from Iran that includes their insistence on the right to continue to enrich uranium and to control the Straits of Hormuz, which we had said were red lines for us before?

22:53Michael Froman:And so those are the kind of issues that need to get sorted out as the vice president heads to Pakistan to try and further these talks. But at the moment, there's just a lot of uncertainty about what has been agreed to and what's included and what's not included.

23:07Dave Kasten:So given that you've participated in talks like this, how do you handicap what the outcome of this even looks like?

23:14Michael Froman:Yeah, look, I think the positive element of it is it was basically just a ceasefire announcement, an agreement to stop bombing of some sort, stop the killing, and to create space for negotiation of the real underlying issues. The real underlying issues remain very difficult to resolve because the Venn diagram between what the U.S. is insisting on and what the Iranians are insisting on, there's almost no overlap. And so that's the really hard part. And the question is whether they can use this two-week period or whatever period it's extended to over time to get to those more difficult issues. It could end up just being a stalemate where there isn't any overall agreement, where the U.S.

23:55Michael Froman:agrees to stop the military action. Iran agrees to open the straits, hopefully without limiting by the Iranian military how many ships can go through and charging a toll and making it a creating more friction in the global trading system and the global energy system. And then just the underlying issues remain unresolved. And it's kind of a frozen conflict that the parties return to over time. And sometimes it'll break out into violence and sometimes there'll be relative periods of peace. Michael, you don't sound that negative on the whole process. And while I was listening to you and I'd seen these other editorials all over the place about if we don't achieve these objectives, we definitely haven't gotten what we wanted.

24:41Michael Froman:And right now, Iran seems to have gotten some of the things it wanted. But then I finally came across Holman Jenkins. His point is the war's not over. Washington has advanced its aims more than Tehran. And the ceasefire isn't permanent. And the night is still young. And our assets are still there. But that just begs the question, what comes next? Whether it's boots on the ground in Karga Island or whether it's boots on the ground somewhere, whether it's renewed attacks on power and infrastructure. But you seem to think we've made some progress and something could be positive to come out of this if it's handled right from here on out or no.

25:22Michael Froman:Well, Joe, you know, Bridge Colby, senior Pentagon official, came to the Council on Foreign Relations at the beginning of the war, laid out the military objectives. many of those military objectives have been achieved or largely achieved in terms of sinking the official navy, damaging their missile and capability, and greatly weakening the Iranian regime. But I think what the U.S. may have underestimated is how much pain the Iranians are willing to go through. They went through an eight-year war with Iraq. And even though it's a weakened regime, two things I think are really important that come out of this period of time.

25:55Michael Froman:One is we've seen the impact of asymmetric warfare with a relatively small number of missiles and drones. Iran can wreak havoc on shipping in the Strait of Hormuz and hold the rest of the global economy hostage. And secondly, with regard to the Strait of Hormuz, Iran could always have closed that. But it had rarely been done or never really been done in this way. And they have learned the lessons of choke points, which is that the U.S. used the dollar and dollar-based institutions over the year to cut off access to the banking system. The Chinese have learned how to use supply chains to cut off access to critical minerals and rare earths.

26:34Michael Froman:And now the Iranians have learned that they have this incredible source of leverage over the rest of the global economy, including the United States, by cutting off the shipping of the Strait of Hormuz. So we leave this conflict or we get to this next phase of this conflict with the Iranians much weakened in some ways, but now having a new tool that they've taken out of the toolbox and demonstrated great leverage with. So what would you advise the United States to do to address that? I think we ought to get back to the fundamental issues, which is what's the status of the Iranian nuclear program?

27:09Michael Froman:We haven't talked about the nearly thousand pounds of near weapons grade enriched uranium that is still there. Are we going to let them be able to enrich uranium in the future or not? And what will be done in the future in terms of if they try and rebuild the nuclear program, try to aid their proxies around the rest of the world? Will the U.S. and Israel continue to take military action against them? Those are the issues that need to be discussed and see if there can be some agreement as to what the guardrails are on an Iranian regime going forward. Diplomacy with Iran and believing what they're saying.

27:47Michael Froman:Is there anything else?

27:48Dave Kasten:Do we need allies to come to the table with this? I mean, if this is the world's problem, do we need a united front to handle the Iranians, even if it's with diplomacy?

28:01Michael Froman:Well, Becky, look, on so many of these issues, that's absolutely right. I mean, we would be in a stronger position if we had allies with us. As the president said correctly, most of the oil doesn't come to the United States, goes to Asia and to a certain degree to Europe, but we are not unaffected. And I think the Europeans and our allies and partners have a very strong interest in how this is resolved, and we need to find ways of engaging them in this process as well, in our position as well.

28:27Dave Kasten:Do you think that's possible, given how this started and what's happened with our relationship?

28:32Michael Froman:You know, we haven't made it any easier by starting, without consulting them and creating, you know, we've had a year of difficult relations with a lot of our allies over tariffs and other issues, Greenland, et cetera. So we haven't made it easy for ourselves. At the end of the day, however, I still believe in national interests. And even though our allies and partners may not love the way they've been treated, they have national interests that they need to follow and hopefully they'll work with us to address them. Great. Ambassador, thank you. Appreciate it.

29:02Dave Kasten:Coming up on Squawk Pod, we are digging into Anthropik's new artificial intelligence model, Mythos, that's alarming software developers and policymakers with its capabilities, particularly in cybersecurity. We'll talk to the head of policy at Palisade Research, Dave Kasten. What comes next? The race isn't over. Everyone's still racing as hard as they can, and they're going to cut as many corners as they can to catch up. And frankly, I think you should ask yourself the question of, would OpenAI have announced today that they were going to do this sort of staggered disclosure thing if it wasn't for the case that Anthropic had led first?

29:42Dave Kasten:Hello, I'm Ozempic. And I'm other GLP ones, kind of like him. Are you shaking a maraca? Nope. I'm shaking the pill version of Ozempic, which no one should ever do except in ads like this.

29:52Michael Froman:A nice disclaimer. Hey, thanks. Ask your doctor about which FDA-approved uses of the Ozempic pen or pill may be right for you. Call 1-833-OZEMPIC or visit Ozempic.com to view the medication guide and learn more about Ozempic semaglutide tablets, 9 mg, and Ozempic semaglutide injection, 2 mg.

30:09Dave Kasten:There's a pill version of Ozempic! Never bet against American grit or American energy. Through innovation, Venture Global is not only building some of the largest energy facilities in the world right here in the United States, but delivering American energy at a fraction of the cost and a fraction of the time. So while others are busy talking, we're busy building. That's Venture Global. That's unstoppable energy.

30:42Dave Kasten:Hey, Mom, now that I have an after-school job, I think I want to buy a drum kit. Okay, Rockstar. Let's look at your account. When kids start making money, Chase has easy-to-use tools and expert bankers to help them learn how to manage it. And next month, I'll buy the bass drum and the cymbals. Oh, and maybe some noise-canceling headphones for you. Aw, you've thought of everything. Help kids save, budget, and build financial independence all with Chase. That's good for kids and good for parents. Visit your local Chase branch and get started. Account subject to approval. JPMorgan Chase Bank, NAMember FDIC.

31:15Dave Kasten:This is Squawk Pod. Straight up on Becky. Three, two, one, cue. And you're watching Squawk Box right here on CNBC. I'm Becky Quick along with Joe Kernan and Andrew Ross Sorkin. Anthropic is starting a controlled release now of its newest frontier model. It's called Mythos. And the model has capabilities that sparked a lot of concern among the public. within Anthropic itself. Joining us right now to talk about it is Dave Kasten. He's the head of policy at Palisade, an AI research firm. The idea that they've built this model, they're not going to release it to the public because they think it's too dangerous, but they're effectively going to give it to other companies, at least ahead of giving it to the public.

31:58Dave Kasten:How is this going to work and what do you think about this? Hi, Andrew. Great to be here. So the proposal that Anthropic has put forwards is what they call Project Glasswing. And their idea is that they're going to partner, at least initially, with 11 other companies that are significant in sort of the technology ecosystem, companies like Amazon, Apple, Broadcom, NVIDIA, that kind of thing, as well as Google, one of their competitors, in order to be able to try to pluck as many of the low-hanging fruit of cybersecurity risks as possible. Anthropic believes that the AI models that they've, the new AI model that they've developed is going to be able to identify a lot of bugs that previous models weren't able to do so effectively and humans weren't able to do so cost effectively.

32:41Dave Kasten:And so they're really hoping to try to clear out the weeds of as many of those risks before other models get to the same place. But what does it mean to every other company out there? So let's say they do this, they get to those companies first, then they release this to the public, and then it's just a field day and a race between the hackers and the companies that weren't in the original 11? Yeah, so I think my suspicion is that the companies that are in the original 11 won't just be working on their own behalf. When you sort of look at the software ecosystems that these companies have, they rely upon a lot of common libraries, whether those made by common third-party vendors or open-source libraries.

33:22Dave Kasten:And so as a result, I suspect much of what Amazon, by example, would be doing, would be looking at open-source libraries that many companies rely on. But they might particularly rely on identifying areas of concern and then trying to get them patched before these sort of bug finding capabilities become more available generally to sort of malicious actors that might want to use them. Is your sense that once that the model itself is obviously not just being used to find bugs and exploit hacking opportunities? I assume it is that much more powerful for the rest of us. No. Yeah, that's right. And in fact, Anthropic has said explicitly that this model wasn't developed for the purpose of being a cybersecurity model.

34:03Dave Kasten:Anthropic, as well as all the other major AI companies, have as their core business goal building automated AI researchers. That's AI models that help them make better AI models, that help them make better AI models, and so on. And so software engineering tasks, particularly the ability to do software engineering tasks on your own, are something that's useful for all kinds of economically valuable things. It just happens that it's also security relevant in that you can use it for cybersecurity, offensive and defensive tasks really easily as well. And, you know, we were talking about this as this most advanced model.

34:36Dave Kasten:You know, where do you think ChatGPT lives in this, meaning, you know, they're working on another model? Does their model have the same issues all over again? I assume Gemini is trying to do the same thing. And how ahead of the game is Anthropic then? Yeah, I think your expectation should be that Anthropic is a little ahead, but not overwhelmingly ahead, and they don't necessarily have much of a permanent moat here. So there was a report from Axios just a couple minutes ago that ChatGPT is also going to have a new model, which is currently codenamed SPUD, that is going to go through a similar process in terms of their sort of slow rollout of capabilities because of similar cybersecurity concerns.

35:19Dave Kasten:I would expect that Gemini is not far behind. Google has certainly been investing in automated AI-based vulnerability finding for quite some time. But the fact that Google is partnering with them probably implies to me that at least for the next couple of months, Anthropic has a little bit of an advantage here with this particular model. Let me ask you just about that advantage, because it really does feel over the last couple of even couple of months now, I'd say last two months, that there was a real move towards Anthropic and its clawed model for a lot of folks in the valley and much even beyond.

35:56Dave Kasten:And the question is, how sustainable do you think that is? Or do you think we're going to be sort of switching back and forth between models because the next ChatGPT or the next Gemini is going to leapfrog wherever they are? Yeah, so I think it's going to be a bit of both. And when I talk to sort of professional software engineers that I know, you tend to see sort of two types of folks. One kind of likes finding a model that's pretty darn good and sticking with sort of that family of models as it advances. The other is kind of constantly switching out at sort of the technical level for the latest and greatest model, whatever that happens to be on a given week.

36:32Dave Kasten:I think you would expect that over time we're going to see more sort of speciation when it comes to the different AI companies offerings. And so you might see more vendor lock-in over time that would occur. I certainly think it's true that when you talk to many coders, they have an expressed preference for either the ChatGPT codex models or more commonly for Claude Code. And I suspect that that sort of taste-based preference might last for some time. So do you have any sense in terms of either declaring a winner or declaring, do you believe it's a multi-winner game? I mean, the other thing that just happened in the last 24 hours is you had Meta releasing its latest new model.

37:11Yep.

37:12Dave Kasten:So, you know, I think the news is still coming in on the meta model in terms of how good it is. Obviously, they released sort of their benchmarking results, but it takes a while for people to actually use the models to sort of say, hey, are those benchmarks really reflecting reality or do they sort of overfit on the benchmarks? Too soon for us to say for sure. It's certainly a good model, whether it's certainly also not as far as anyone can tell a model that beats OpenAI or Anthropic, but certainly a credible showing by meta. I think you would expect to see that the race isn't over. Everyone's still racing as hard as they can, and they're going to cut as many corners as they can to catch up.

37:48Dave Kasten:And frankly, I think you should ask yourself the question of would OpenAI have announced today that they were going to do this sort of staggered disclosure thing if it wasn't for the case that Anthropic had led first? I think some people might be skeptical of that, including myself. Okay, Dave, we got to run, but my final question for you is, you just said cutting corners. Usually when you hear that phrase, at least my ears perk up, what does cutting corners mean in this context? Yeah, so I think something that non-experts don't know is that every AI model sometimes does bad behavior. It doesn't do what users want.

38:22Dave Kasten:It sometimes lies to users. It sometimes engages in all kinds of blistered behavior. You can read Anthropics disclosures on this as well as opening eyes on everyone else's. the more that you're trying to hit a timeline for release of a new model, or you're trying to hit a given benchmark number, the more that you might be willing to accept some occasional bad behavior in order to crank more performance out of your model. And so I think one thing that we at Palisade are worried about is that people will have a lot of incentives to cut corners on safety in order to either get back in the race in the case of places like Meta, or to stay in the race in the case of places like OpenAI, Google, or Anthropoc.

38:54Dave Kasten:Thank you, Dave, for scaring us a little bit this morning. Appreciate it. Thanks so much. That is Squawk Pod for today, this Thursday. Thanks for listening. Squawk Box is hosted by Joe Kernan, Becky Quick, and Andrew Ross Sorkin. You can tune in weekday mornings on CNBC at 6 Eastern. To get the smartest takes and analysis from our TV show right into your ears, follow Squawk Pod wherever you get your podcasts. We'll meet you right back here tomorrow.

39:25Michael Froman:We are clear. Thanks, guys. Thank you so much.

39:39Dave Kasten:We'll be right back.

40:00Dave Kasten:and good for parents. Visit your local Chase branch and get started. Account subject to approval. JPMorgan Chase Bank, NA member, FDIC.