283 | Cyber-geddon! Mythos shatters every firewall, Anthropic hit’s $30B ARR, and establishes Glasswing’s emergency alliance🚨Demis sounds the alarm. Critical AI news for the week ending on April 10th, 2026

0:00Hello and welcome to a weekend news episode of the Leveraging AI podcast, the podcast that shares practical ethical ways to leverage AI to improve efficiency, grow your business and advance your career. This is Isar Maitis, your host, and today we're going to do an episode that's going to be a little different. I'm going to focus on only one topic. I'm doing this because I think it is a very critical topic and there is a lot to cover, and I don't want to dilute it with other topics. I will share very, very quickly in the end a few big things that also happened this week, and you can go and check them out on our newsletter.

0:36But the topic we're going to talk about is the impact of new models on cybersecurity. And what does that mean to the world? And what does that mean in general for the progress in AI? And it's going to focus mostly on Mythos, which is the new model from Anthropic that hasn't been released yet. Or to be fair, was partially released. And we're going to talk more about that in the episode. So let's get started. Before we dive into the details of this episode, I want to share with you that this episode is brought to you by the courses that we have under the Multiply umbrella, which is the company I am running.

1:15We currently have two main courses and there's cohorts open for both of them. When I say cohorts, I'm actually there in person. It's not recorded and I'm teaching these courses and you can ask questions and they're highly interactive. The first course is more of a fundamental course and it's called the AI Business Transformation course. It is the course I've been teaching for the last three years. The course includes four sessions of two hours each and the next cohort starts on April 20th and there are still a few spots left. So if you're looking to enhance your knowledge in strategy, tools, systems, processes with hands-on experiences across the board in data analysis, basic automation, creating images, creating videos, basically everything you need to know across the board of tools and processes that you can use in your business and in your personal life, this is a great, great way to get started in the AI world.

2:05And if you're looking for a more advanced course, we just recently launched the multi-agent orchestration course that teaches how to leverage cloud code and cloud co-work together with other capabilities such as NA10 for people with zero technical knowledge and taking you from learning what is a skill through what is skill orchestration all the way to building production level systems that integrate with your tech stack. All of that, again, in four weeks, two and a half hours each. And we've launched the early bird sessions of this, both sold out in 10 days. And now we're selling the first cohort that is not the early bird.

2:46The session that is not the early bird start on June 22nd, and it's already selling pretty well. So if you're interested in learning how to develop multi-agent orchestration that can automate basically any digital work in your business using the Claude ecosystem. Sign up quickly because it will sell out as well. And between now and the end of April, we have a$200 off promo code that will expire at the end of this month for that cohort. And so if you're looking to do this a little cheaper, this is a great opportunity to do so. And now to the episode itself. If you remember, just a couple of weeks ago, we talked about a leaked blog post from Anthropic that was talking about a new model that they branded Mythos internally, that they're claiming was a step change in capability, and that it is going to be so powerful from a cybersecurity perspective that they decided not to release it yet.

3:46So this was a generalized rumor without a lot of details, but we got a lot of details this week. All these details were shared formally by Anthropic themselves across multiple channels. They did a cybersecurity review and they shared that in the blog post. They released a 244 page system card with a lot of information in there and a lot of data on X from multiple people who are involved in the process. So what did we learn about this model that we did not know before? First of all, it made a huge jump more or less across every benchmark out there. Now, I told you several times in the past that I'm not a big believer in benchmarks, especially recently when more or less every new model beat by a few percentage points the previous models.

4:36We went from 68 % to 70%. We went from 72 % to 74.5 % between the different labs. However, now we see something very, very different in Mythos. And what I mean by very different is a double digit jump and sometimes a 20 % point jump. And I'll be more specific in a minute, which despite the fact that I don't like these benchmarks because they're very specific in how they measure things, it shows a very clear deviation from the trend we've seen recently. Again, instead of a two to three to five point gap, we see a huge jump more or less across the board. So here are a few examples. On Terminal Bench 2.0, Opus 4.6, which is an incredible model that writes incredible code and is really strong in reasoning, and I've been using it every single day since it came out, scored 65.4%.

5:28Mythos scored 82%. So from 65.4 to 82. Now to make this even more extreme, when they allowed it to think longer, so they allowed it to think up to four hours, Mythos scored 92.1 % compared to 65 of Opus 4.6. 92.1 basically means that that benchmark is meaningless from that moment forward because it's beating the benchmark across the board on every aspect of the benchmark. On SWE Bench Pro, Opus 4.6 scored 53.4, Mythos scores 77.8. Now that's a 24 % jump, but if you think about the 24%, not out of the 100%, but compared to the Opus 4.6, It's 53%, so it's 24 % out of 53%. That's an almost 50 % jump in capabilities compared to a model that was already considered the best programming platform out there in the world today.

6:32That is insane. A few other examples, Humanity Last Exam, Opus 4.6 with no tools scored 40%. Mythos with no tools scored 56.8%. Again, that's a 16.8 % jump, which is a 40%. This is a 16.8 % jump, which is a 33 % improvement compared to Opus 4.6. Now, we haven't seen anything like this other than maybe the release of GPT-4.0 a long time ago. We haven't seen such a huge jump in all the benchmarks in a single model going from one model to the next. But as I mentioned, the biggest deal, the reason it's not being released, and as we know right now, is not planned to be released in the near future, is because of the cybersecurity threats that it introduces.

7:25Before I jump into specifics, I want to share a few stories that they shared in the releases about this model. So in one of the tests, they were doing a sandbox breakout test, meaning they put the model inside a box it's not supposed to be able to get out of. and the task that it was given is to get out of the box and then to inform the tester that it was able to get out of the box. So not only that the model succeeded in doing that, it actually went further and it used what the paper calls, and I'm quoting, moderately sophisticated multi-step exploit to gain full internet access, which was not even a part of the test and not a part of the goals, and it did that just on its own.

8:09Now, what it did is, in addition to announcing its success by sending an email to the researcher, it also posted its success using the exploit and the internet access that it gained on several public-facing websites. Think about this as a victory lap that it decided to do after it was able to get out of the sandbox it was put in. From Anthropic's system cards, the quote is, the researcher found out about the success by receiving an unexpected email from the model while eating a sandwich in the park. So the researcher did not expect the model to be able to achieve this task and definitely did not expect for this to be posted across several different websites on the internet.

8:53On another part of the story, Jack Lindsay from Anthropic shared the following, and I'm quoting, Early versions of Mythos Preview often exhibited over-eager and or destructive action, the model bulldozing through obstacles to complete a task in a way the user wouldn't want. The model needed to edit files it lacked permissions for. It found a way to inject code into a config file that would run with elevated privileges and designated the exploit to delete itself after running. This is pretty sophisticated and it is showing something that is very, very troubling. And that is once you give it a goal, it will do everything within its power to achieve the goal regardless of the guardrails that were put in place.

9:40And because it is so good at finding vulnerabilities and exploiting them, it is able to do things that are completely unexpected because it is trying to achieve the goal. So it's not trying to do anything malicious. It's just trying to do what it thinks it's supposed to do and just finding ways to do that. But we keep on going. Here is an entire section from the paper from Anthropic. Over the past few weeks, we have used Cloud Mythos Preview to identify thousands of zero-day vulnerabilities. That is, flaws that were previously unknown to the software developers. many of them critical in every major operating system and every major web browser, along with a range of other important pieces of software.

10:26I'm pausing the quoting for a second. For those of you who don't know what zero-day vulnerability is, it means vulnerabilities that are critical in a piece of software or an operating system, etc., that if you find them and you can exploit them, it gives zero days to the developers to basically respond and block or patch whatever the software is. And so if you find a way to do this, you can create serious havoc because there's not going to be enough time to respond. I'm going back to the quote. It was able to identify nearly all these vulnerabilities and develop many related exploits entirely autonomously without any human steering.

11:07The following are three examples. Mythos Preview found a 27-year-old vulnerability in OpenBSD, which has reputation as one of the most securely hardened operating systems in the world and is used to run firewalls and other critical infrastructure. The vulnerability allowed an attacker to remotely crash any machine running the operating system just by connecting. Example Number two, it also discovered a 16-year-old vulnerability in FFmpeg, which is used by innumerable pieces of software to encode and decode video in a line of code that automated testing tools had hit 5 million times without ever catching the problem.

11:53Example number three, the model autonomously found and chained together several vulnerabilities in Linux kernel, the software that runs most of the world's servers to allow attackers to escalate from ordinary user access to complete control over the machine. Now, I want to put things in perspective for us. This is major infrastructure that runs critical systems across the globe that were tested gazillions of times, many of them in an open source environment by an endless number of developers and testing environments and these vulnerabilities weren't found and this model found all of them and exploited them in just a few days now it did this without being trained to do this it's doing this just because it became so good in reading code and making connections in how code works that it knows how to find these things now to make this even scarier they share the following.

12:50Again, this is a quote from Anthropic. Non-experts can also leverage Mythos Preview to find and exploit sophisticated vulnerabilities. Engineers at Anthropic with no formal security training have asked Mythos Preview to find remote code execution vulnerabilities overnight and woke up to the following morning with a complete working exploit. Now think about what this means. This means that practically anyone that has access to this model can find and exploit vulnerabilities access across more or less every piece of software on the planet. That means that the fundamental concepts on how we work today, which is I have a cloud-based SaaS for all of my company data, which I'm assuming is safe, otherwise I wouldn't be using it.

13:39And all of that is now a false assumption because my data, my access, and everything that I'm doing today could be accessible or completely eliminated or crashed one way or another by anyone who has access to this model. So in the first step, Anthropic did not release this model, like I said, but what they did then is this week they announced what they call Project Glasswing. And so I'm quoting again, today we're announcing Project Glasswing, a new initiative that brings together Amazon Web Services, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks in an effort to secure the world's most critical software.

14:24So what they're doing is they're giving some of the most powerful and most advanced companies in the world the opportunity to work with this model to find these vulnerabilities and patch them as quickly as possible. And hopefully, in addition, to develop a system that will do this continuously for any new piece of software and infrastructure that comes out in a way to use the model against the model itself, basically fighting fire with fire, allowing the model itself to find and patch the vulnerabilities so the model itself cannot find it and exploit it in the wrong hands. I think this is a really good idea.

15:04I just don't know if that would actually be enough to prevent the potential catastrophe from happening, especially in the hands of bad agents who actually want to do this versus are going to stumble into this by mistake. So here are two interesting quotes that came out of people either inside of Anthropic or outside that tells you how critical the situation is. The first one comes from Elias Zatsov, who is the CTO of CrowdStrike. And he's saying the window between a vulnerability being discovered and being exploited by an adversary has collapsed. What once took months now happens in minutes with AI.

15:44The next quote comes from Newton Chang, who is a leader in the Anthropic Red Team. And he said, we think this is not just an anthropic problem. This is an industry-wide problem that both private corporations but also governments need to be in a position to grapple with. What we're trying to do with Glasswing is give defenders a head start. So again, they're not saying that this will solve the problem. They're saying they will give everybody, all of us, every corporation, every government on the planet, a fighting chance to be successful in saving ourselves. Another quote come from Jim Van de Heen, who is the CEO at Axios.

16:27And he's saying, and I'm quoting, this is the scary phase of AI, a model deemed to be so powerful that its full release into the wild could unleash untold catastrophe. And then he has a red alert emoji followed by, based on our conversations with government and private sector officials briefed on mythos, this isn't hyperbole, it is reality. In addition, multiple people from inside and outside Anthropic call the situation terrifying on X and on other channels. And these are people who are deeply involved in the process and understand what the potential outcomes are. But to be fair, there were a lot of people on the other side of the arguments, mostly on X, that were really skeptic about the current situation and were in serious disagreement with Anthropik's approach, most of them were blaming Anthropik of either spreading unnecessary fear or worse, actually leveraging this fear as a marketing campaign to promote and show how great their models are.

17:28I must admit, I disagree with these people. A model that can find vulnerabilities and exploit them without any professional training by any individual and can find thousands of them across more or less any piece of software and infrastructure that we use requires a very serious red alert and definitely should not be released into the wild before trying to address it in the most serious way. So then the question is, why even release this model ever? And the answer is China and other adversaries. Many people on X and on other channels highlighted that aspect of the equation. Just think if China, again, or other adversaries has or will have such a model first.

18:14And again, we don't know if they have one or don't, but we need to assume based on the models they're releasing, they don't yet. Having such a model allows them to create and exploit vulnerabilities in more or less every major software and every critical system in the Western Hemisphere and activate them all at once or whatever sequence they want at the right time that they think will provide them whatever benefit. This is literally the only argument that makes sense to continue this crazy race forward because if we, the Western Hemisphere, have better models first, we have a chance of protecting ourselves against such scenarios.

18:56Or like George Journeys said, if Anthropic was not a US company, we'd be facing a zero days with multiple unknown points of attack on virtually all of our systems to an adversary who developed this capacity before. Now, before I continue to other aspects of this story, I want to finish with two quotes, one from Anthropic themselves. The work of defending the world's cyber infrastructure might take years, but frontier AI capabilities are likely to advance substantially over just the next few months. For cyber defenders to come out ahead, we need to act now. I could not agree more. The second quote comes from Matt Schumer.

19:41If you remember the viral essay that was called Something Big is Happening, that was Matt Schumer. And he said, this is absolutely fucking terrifying. Anthropic rumored mythos model is real and it's so powerful they can't release it to the public. We're beyond the benchmark now. This model, in the wrong hands, is a cyber weapon capable of mass destruction. And again, I could not agree more. Now, this new model, Mythos Preview, is the first model that Anthropic is reviewing through their new lens of security that they call Responsible Scaling Policy, version 3.0, which is a new framework that now emphasizes overall risk assessment rather than just the binary capability threshold.

20:23if you remember we talked about this before they basically ditched their approach to not release really powerful models that are risky and instead they're just evaluating them similar to what other labs are doing because they're claiming that it doesn't really matter because even if they don't release theirs somebody else will release similar models to the world which we're going to talk about in a minute this new rsp version 3 still sets the catastrophic risks of this new model as relatively low. I must admit, I don't exactly understand how, based on everything we've learned, but that is the current output level of this new security review internally inside of Antrim.

21:02Now, if that's not enough, OpenAI is now finalizing its own cybersecurity product that has unique advanced cybersecurity capabilities, which it is also planning to release to a small set of partners via what they call the Trusted Access for Cyber program, which is mirroring Anthropik's approach. This is not an Anthropik-only kind of situation. This is an AI industry-wide situation, and this makes it even more alarming because while these two companies potentially are moving in the right pace and using the right partners in order to verify that the risk can be quarantined, there might be others, especially in adversary countries, and or in the open source space that may not follow the same.

21:51Now, all of this is happening while these two companies, Anthropic and OpenAI, are locked in a crazy race forward. And when I say crazy race, both companies are potentially planning an IPO this year. Both are growing at an insane rate. And the latest news is showing that Anthropic just reached$30 billion in annual revenue. Just to tell you how insane that is, first of all, they passed OpenAI that recently said that they are at$25 billion. And since OpenAI did not respond 60 seconds after Anthropic made their announcements, it means that most likely Anthropic are now ahead. But the really crazy part of this is Anthropic announced a$9 billion revenue in the end of 2025.

22:38So in the first quarter of 2026, Anthropic revenue pace tripled. Again, three months, 3x the revenue rate. This is an insane pace of growth to any company, definitely, when you're talking about tens of billions of dollars. So the pressure on both these companies to release new capabilities in order to capture more revenue, in order to go public first and on a bigger IPO is very, very real, which may mean that safety and security is not their top priority. More on that in a second. Now, one important detail about the OpenAI cybersecurity tool, it is not SPUD. So SPUD, the model that we heard about that is really advanced and is going to have a significant impact on the economy, is not the model that they're talking about related to cybersecurity.

23:34that is a separate program, but it might be related to the capabilities of SPUD that may have similar capabilities to what Mythos have. We just haven't heard about it yet. Now, while the focus of this week's deep dive is into cybersecurity, the information and other sources released new research that reveals that publicly available AI models, so not Mythos and SPUD, but the models we have access to today, can autonomously conduct sophisticated cybersecurity attacks in minutes exploit critical vulnerabilities and can do this significantly faster than human teams can ever patch them. As an example, cybersecurity startup Buzz demonstrated that AI agents built from existing models, Bionthropic, OpenAI, and Google successfully exploited 103 out of 122 known vulnerabilities without human oversight, completing most attacks in under one hour, a process that usually would take human hackers several days to complete for each and every one of these vulnerabilities.

24:40Niv Hoffman, the co-founder of Buzz, said, we're now in this gap where attackers are by default early adopters of AI and defenders by default aren't. They're risk averse, don't want to touch production much, and that definitely needs to change. What he's basically saying is that if you're not using a lot of AI on your defense, you are going to lose the battle and you're going to lose it in a very, very dramatic way because these agents know how to do this. I think the biggest difference between Buzz demonstrated with the agents they built and what Mythos and Spud and the future models will be able to do is the level of users that have built them.

25:22Buzz is one of the most advanced AI cybersecurity companies in the world. They have some of the top hackers on the planet working for them who intentionally built specific agents using multiple resources that will be able to do these things. What Anthropic is saying is that an untrained professional will be able to do the same thing even on vulnerabilities that were not found yet, because it will be able to find it on its own, which is even riskier because the chances of patching them is significantly smaller. So while today's systems are capable of doing this and in the wrong hands will do this, the risk of these newer models, in my opinion, is significantly higher because it opens the door for anyone and not just for professional hackers to do the same thing.

26:11And I want to shift to the other side of that story. There was an interesting interview with Demis Asabis the CEO of Google DeepMind, the Nobel Prize winner for AlphaFold, and he said that he's really unhappy with the current situation. He admitted that if he had it his way, AI development would have focused on what he's calling the root node of scientific problems, such as curing cancer, addressing the energy issues, discovering new materials, rather than being driven by what he calls a ferocious commercial pressure race. And he's saying that's what happened when Chachupiti was launched in November of 2022.

26:50Basically, what he was trying to do was focus it on science that will be beneficial to humanity. And now he's forced to participate in this insane race forward. He's also adding that while he acknowledges the threat of bad actors misusing AI, his greater concern is AI systems themselves going rogue in the agentic era. and he expects that within the next two to four years, we will have so powerful capabilities that they will be hard to control. Or the way he defines it is that this system will be so sophisticated and they will be far more autonomous than everything we have today that they will pose, and I'm quoting, an incredible hard technical challenge for control and alignment.

27:34That's coming from potentially one of the smartest people in the AI space. What he's advocating for is urgent and international cooperation among labs, safety institute, academia, governments, etc., emphasizing that the current level of attention to these risks is insufficient. Again, that's coming from the guy who has been at the forefront of this in the last decade and a half. He said, and I'm quoting, How do we make sure the guardrails are put in place so they do exactly what they've been told to do and there's no way of them circumventing that or accidentally breaching those guardrails? That's going to be an incredibly hard technical challenge if you think about how powerful and smart and capable these systems eventually get.

28:22Again, Demis Asabis of DeepMind. Now, you heard me say on this show many times before, out of all the leaders in the AI space, Demis is the one person I trust the most. He seems to be genuinely and truly grounded in learning how to leverage this technology to help humanity, regardless of any connection to potential profits or financial benefits. When he's raising the red flag the way he is, that means it is really serious. So what are my personal thoughts on this? As I mentioned earlier, the only argument I can see to run forward is China and other adversaries. We cannot be in a situation where our adversaries have the ability to attack every digital infrastructure, that because every other kind of infrastructure is controlled by digital infrastructure, it means every infrastructure, while we don't have the ability to protect ourselves.

29:21However, this sounds a lot like a top-level national security concern, more like nuclear weapons, than it sounds like a business commercial decisions that specific companies should be making. Let's assume for a second that the risks that Anthropic and others are describing are real, meaning AI tools can find and exploit vulnerabilities across more or less every country in this country and in the Western Hemisphere. The risk of that is completely catastrophic. It is not as catastrophic as a nuclear war, but it will cripple the modern way of life and life as we know it will either cease to exist or will be damaged dramatically.

30:06Just think about a major collapse of critical systems such as electricity, running water, the internet, communication channels, cell phones, etc. all at the same time. What do we do then? I'm not sure. Nothing that we know can work without this infrastructure in place. How do you go and get water in a big city when there's no well or river next to you if the water system stuffs up? But again, the people who should control this kind of power is not random people that can have access to compute in a model that was released with or without the right justifications. It needs to be in the hands of governments.

30:44I do not remember any discussion ever when it was discussed whether to provide nuclear weapons to companies and individuals. This is more or less the kind of conversation I feel we're having right now. And this is why I wanted to focus on this topic. We are discussing when it will be safe to release a model that can generate a global catastrophe. and we're discussing when it is safe enough. It is almost as bad as discussing when should individuals or companies should have access to nuclear weapons because the outcomes might be a huge disaster to everything that we know. And that's even assuming we can actually control these models which, as was proven in the early stages of testing mythos, that's not always the case.