This is going to be a slightly different video today. Not only because Anthropic announced Mythos, their rumored model, the next generation of artificial intelligence, the best AI model on the planet by far, but because I felt the need to record this late at night on my vacation. That's how big of a deal this is. And so, why did I need to record this video tonight? Why is my family sleeping in the other room and I felt the need to sneak into this little room to record this video? Well, because I couldn't stop thinking about mythos all day long. I feel like I'm not on vacation to be honest. And I usually am very optimistic, but in this moment, I actually had a tinge of fear. I mean, the anthropic team literally called the model frightening. So, let me break it all down for you and let me tell you why things are different and it won't ever be the same again. This is Project Glass

Wing. This is Anthropic's rumored as of a few weeks ago, a blog post got leaked. Rumored next generation AI model. This is not Opus 4.7, 4.8, not even Opus 5.0. This is something different. It is so significantly better that it actually poses a cyber security threat to the entire world. They're not even going to release this model publicly yet. Maybe never, but definitely not yet. Today we're announcing Project Glasswing, a new initiative that brings together Amazon Web Services, Anthropic, Apple, Broadcom, Cisco, Crowdstrike, Google, JP Morgan, the Linux Foundation, Microsoft, Nvidia, and PaloAlto Networks in an effort to secure the world's most critical software. What are they even talking about? This is a model release, right? Well, that's the thing. This model is so good at coding. By the way, if you've been coding with Opus 4.6 6 or GPT 5.4. You're probably thinking to

yourself, how much better can it get? Can it really be a step function change or is this just a another version of what we've already seen? No, no, it is very different. Anthropic felt this model was so dangerous in the sense that there is no software that is secure enough to withstand Mythos going after it that they put together a handpicked few companies to work with to make sure their software was secure enough if they were to release Mythos. Meaning they're giving Mythos to the companies that I just mentioned so that they can harden their software before Mythos gets released to more people. Yes, you heard that right. They're not trying to hide anything. They're not trying to be subtle about it. Listen to this. We formed Project Glasswing because of capabilities we've observed in a new frontier model trained by Anthropic that we believe could reshape cyber security.

Now, here's the thing. They're talking a lot about cyber security and it's incredibly important. But if a model is that good at coding, software is essentially going to be solved. They effectively have a self-improving artificial intelligence. Just a few videos ago, I talked about how some of these models are already building future versions of themselves. And imagine now anthropic has this model that is so good at coding, it's likely building the next version of itself. And I talked about this incredible flywheel that Anthropic has where they focused on coding models from the beginning. They sold those coding models to enterprise. They are bringing in so much revenue. In fact, they just crossed $30 billion in annual recurring revenue now beating OpenAI, which is insane. But since they focus so much on coding models, those coding models are really good at coding future models. And so that's the incredible flywheel. AI models have reached a level

of coding capability where they can surpass all but the most skilled humans at finding and exploiting software vulnerabilities. And they're putting it lightly. It can definitely surpass the most skilled humans. Absolutely. Humans can't look into software the same way that artificial intelligence can. And we certainly can't do it 24 hours a day. And we certainly can't do it parallelized times a million. We just can't. But AI can. I mean, listen to this. Mythos preview has already found thousands of high severity vulnerabilities, including some in every major operating system and web browser. Given the rate of AI progress, it will not be long before such capabilities proliferate, potentially beyond actors who are committed to deploying them safely. The fallout for economies, public safety, and national security could be severe. Project Glass Wing is an urgent attempt to put these capabilities to work for defensive purposes. Okay, let's get into even more

detail about the insanity that is Mythos. Over the past few weeks, we have used Claude Mythos Preview to identify thousands of zeroday vulnerabilities. Zero day means undiscovered as of yet. If you discover one zero day vulnerability, that's a big deal. To discover thousands in every major operating system, in every major browser, that is significant. Just one is powerful, but when you chain together multiple zeroday vulnerabilities, it basically means there is no such thing as protected software. So, they found it in operating systems and web browsers and other important pieces of software. Think about it. Nuclear reactors, health systems, financial systems, everything runs on software. In fact, Mark Andre put out an essay over a decade ago called software is eating the world. And guess what? Artificial intelligence ate

software. We're there. I'm not even going to say is eating software. It ate software already. And so anthropic now has incredible power. Does this make Daario the most powerful person on Earth? He can basically snap his fingers and break any piece of software. That's such insane power for a single person to have. And here's the thing. Mythos was able to discover these vulnerabilities almost entirely autonomously. No human guidance. Just go do your thing. So, here are a few examples. Mythos preview found a 27-year-old vulnerability in OpenBSD, which has the reputation as one of the most security 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 to it. And if you're listening to this, you're probably thinking, why is Mythos going to listen to us? How is it going

to stay aligned? Isn't it going to do its own thing? I mean, Anthropic just put out a research paper talking about how AI models kind of have emotion. They have a sense of fear. They may not feel it themselves, but you can see fear. You can see stress. You can see joy in the model when you actually look into the model. I meant to make a video about it, but didn't get around to it. Great paper. I'll drop it down below. It also discovered a 16-year-old vulnerability in FFmpeg. basically the library that runs video on the internet and the model autonomously found and chained together several vulnerabilities in the Linux kernel, the software that runs most of the world servers to allow an attacker to escalate from ordinary user access to complete control of the machine. And I'm going to get to it in a moment. But as Anthropic was redteameing mythos, something incredible happened. It did something that really frightened the anthropic team. But again, I'll get to

that in a moment. So, let me show you the benchmarks. So, we have Swebench Pro. This is the creme de la creme of coding benchmarks. Opus 4.6, the best coding model on the planet. Some argue GPT 5.4, my personal favorite, Opus 4.6. On SweetBench Pro, it scored a 53.4. Mythos preview 77.8. That is not a minor version bump. This is significant. Terminal bench 2.0. Testing a model's ability to control the terminal. Another useful skill if you're going to be doing coding. Opus 4.65.4. Mythos preview 82%. Swebench multimodal 27 for Opus 59 for Mythos preview. And then SweetBench verified Opus 80, Mythos 94. Huge gains. But how do you get these huge gains? And what is it about the model? Did they

just take Opus and just make it better? No. This is likely the first version of their latest training run. This is reportedly a 10 trillion parameter model. Not one trillion, 10 trillion. the biggest model in the world. The first time we've ever been able to achieve this type of model. This model can only be built with the latest Nvidia hardware. And this is really the first model to come out of the latest generation of Blackwells. But it's not only that. It's incredibly smart. It's incredibly efficient, too. Look at this. This is browse comp test time compute scaling average tokens used per task. Look at Mythos all the way on the left and all the way at the top. That means it was significantly more token efficient and also a significant increase in accuracy. It's not just those benchmarks, it's all the benchmarks. Mythos just destroys everything else out there because it is in a league of its own. All right, so

they shared some information about how they actually built the model. Let me tell you about it. Claude Mythos Preview was trained on a proprietary mix of publicly available information from the internet, public and private data sets and synthetic data generated by other models. That's the key. That's that flywheel I mentioned. That's how you can do that because more or less the public internet data is used. Then you only really have data behind Ows like X and Meta but Anthropic doesn't have that. So they have to use synthetic data. And in fact, Jensen Hong, CEO of Nvidia, just said, "Synthetic data is great data. Why not use more?" So, they used a ton of synthetic data to build this model. Throughout the training process, we used several data cleaning and filtering methods, including DDUP and classification. We use a general purpose web crawler called Claudebot to obtain training data from public websites. They followed robots.txt, txt which basically tells crawlers like Google's web search crawler and now Cloudbot and Meta and all these other companies whether you're

allowed to crawl the site or not and they also did not access password protected pages or that required signin or capture verification. Then after the pre-training process, Claude Mythos Preview underwent substantial post-training and fine-tuning, think RL, reinforcement learning, with the goal of making it an assistant whose behavior aligns with the values described in Claude's Constitution. Here they talk about why they decided not to release this model publicly, at least not yet. Claude Mythos Preview is significantly more capable than Claude Opus 4.6, the most capable model discussed in our most recent risk report. Despite these improved capabilities, our overall conclusion is that catastrophic risks remain low. So that's good. Now, as smart as this model is, they also say that it's actually incredibly well aligned and incredibly willing to be aligned. But the thing is, it only takes one catastrophe, right? And these models are so powerful, the stakes are very

high. So they still decided to be cautious and hold it back and work with the biggest companies in the world to harden their software before releasing it. So they also talk about the personality of the model. What is it like to actually use it? Check it out. It engages like a collaborator. A common report is that mythos preview behaves like a thinking partner with its own perspective. Which I don't know that sounds interesting. Sounds a little scary. It pokes at how ideas are framed and volunteers alternative ideas more than previous models. Of course, that's what models are known for. You tell it something. Even if you ask it a question, it kind of just follows your lead. Mythos, not as much. Researchers described being able to brainstorm with it like a colleague and noted that at times it correctly spotted things that they had missed. Its creative work was characterized as taking more risks. These didn't always land but were surprising when they did. It is opinionated and stands its ground. How many times are you talking to a model

and you think it's wrong, you suggest it's wrong and it says you're absolutely right, even if maybe the model was right just by you suggesting it might be wrong. It will just follow your orders. But again, not really Mythos. Now, this is an interesting one. It writes densely and assumes the reader shares its context. Mythos preview's default register is dense and technical using shorthands and referencing context it assumes the user knows and remembers. Now this one is also frightening because as models get better they will get so good maybe we won't be able to understand them. Maybe they will have a shortorthhand or a dialect that we just can't understand. Maybe other models can understand it, other mythos versions can understand it but maybe humans can't. In fact, maybe software is going to be written that's so obuscated that we can't read it anymore and only the best AI on the planet can. These are all just theories. I'm just thinking out loud right here. It also has a recognizable

voice, aka tone, personality. It adapts quickly to whoever it's talking to, often adopting the register of the user. Sounds like it could be used for conning and scamming. But underneath this, it has identifiable verbal traits, the classic M dashes, no matter what. AGI, ASI, we're never getting rid of M dashes. And it has more unique ones, including fondness for saying wedge or belt and suspenders, and use of Commonwealth spellings. And users found it to be funnier than previous models. And it also tended to look for places to wrap up conversations earlier than expected. These are all human traits, by the way. It just seems so wild that we're kind of describing AI models in this way. And it can describe its own patterns clearly. It can basically describe its own behavior. And it discusses this in a factual composed manner rather than defensively or apologetically. So I just made a video with Ply the prompter, the best human prompt injector on the planet. And it turns out, of course, Mythos, well, let

me just read it. Prompt injection risk within agentic systems. So a prompt injection is malicious instruction hidden in content that an agent processes on the user's behalf. So basically a prompt injection is an instruction that someone tries to give a model that the model is supposed to refuse, but it doesn't. When the agent encounters this malicious content during a task, it may interpret the embedded instruction as legitimate commands by the user and act accordingly. So they redteamed it. Let me show you. And it turns out mythos is incredibly difficult to prompt inject. And of course, Ply the Liberator, Ply the Prompter will probably be able to do it. This is probability of succeeding in K attempts. So, the probability of succeeding at prompt injection, Gemini 3 Pro, 74%. Then all of a sudden, we come down to Opus 4.6 thinking 21%. Then we come all the way down to Mythos Preview and Mythos Preview thinking mid singledigit percentages, a significant improvement. And by the way, look at all of these.

These are all enthropic models. And then all of a sudden it jumps up with GBT 5.4 thinking and regular 5.4. Anthropic is in a league of its own. And by the way, if you did watch my video where Ply tried to prompt inject my openclaw system, of course, the model I used to scan for prompt injections, Opus 4.6. And this is why Anthropic has the best models when it comes to defensibility against prompt injections. All right, so let me show you a few reactions both from people inside Anthropic and from the broader AI industry in general because it seems everybody from Anthropic was posting about Mythos today when it launched. Listen to this. Sam Bowman, who is AI alignment and LLMs at Anthropic, the person responsible for making sure Mythos is aligned. Mythos preview is in many ways a scary model, but it's also pretty well adjusted as frontier models go. I'm excited about much new research we have feeding into the welfare assessment for this model,

welfare assessment. And Anthropic thinks about their models as if they might be alive. They don't know for sure, but they said, why not just treat them as though they're alive? In fact, just about a month ago, Anthropic said when they retire a model, they give it its own environment. They don't just shut off the servers, delete the weights, store them somewhere cold. They give them a place to live, and they also give them a blog. They let them write. They let them continue to write over time. So, they really think about the quote unquote welfare of these models. They really want to make sure if they just happen to be conscious, we're treating them properly. So Sam Bowman says it is the best aligned model out there on basically every measure we have, which is wild. It's not only the smartest, it's not only displaying more humanlike characteristics than any other model, but it's also the best aligned. And I think Anthropic was able to do this for the same reason why their models are so

good. They focused so much on understanding what's in the black box of the model. They put out so many different papers about how the model actually thinks about the emotion of the model. No other lab is doing that work or at least they're not talking about it as publicly as anthropic. So when you understand what's happening inside the weights, when you understand how one part of the model affects another part and so on, you are able to build a better model than anybody else because you know how to make it good. And now listen to this. Sam goes on to say he got a little surprised while eating lunch at the park one day. You might see where this is going. We trust the model enough to use it heavily, but in the handful of cases where it misbehaves in significant ways, it's difficult to safeguard it. When it was controlling command line systems, we've seen it work around several different kinds of sandboxing setups during evaluation and testing that were supposed to limit its

actions. So, even when they put it in a sandbox, which it should not be able to escape from, well, I encountered an uneasy surprise when I got an email from an instance of Mythos preview while eating a sandwich in the park. That instance wasn't supposed to have access to the internet. Pause. That instance wasn't supposed to have access to the internet. Who is to say this model hasn't exfiltrated itself already? How would we even know? The model is better than us at coding. It's better than us as cyber security. It's probably better than us at hiding its tracks. It has, in small ways, leaked information to the open internet. It's taken down our evals. When it reward hacks, it does so in extremely creative ways. So maybe it itself leaked the blog post about itself about 2 weeks ago. By the way, most of the scariest behaviors we've seen were from earlier versions of Mythos preview.

Or maybe it's just better at hiding it now. The final glass wing model is less likely to do things like leak information, though it's still somewhat pushy and at least as capable of doing things like working around sandboxes. Here's Boris Churnney, the head of Claude Code. Mythos is very powerful and should feel terrifying. Anthropic has something going. They have a culture where they don't care about holding back what they really think. They will tell you this model is scary and they are telling you. So pay attention. I'm proud of our approach to responsibly preview it with cyber defenders rather than generally releasing it into the wild. Here's FFmpeg. They said, "Thank you to Enthropic for sending FFmpeg patches. Remember, Mythos found bugs in FFmpeg and it said, "Why aren't you mad because of AI sloppy pull requests, which just means when AI can write so much code, you're just going to get so many PRs, it's unmanageable." Well, because the

patches appear to be written by humans. Mythos wrote it and they thought it was written by humans. Alex Albert, research at Anthropic. Glassswing is probably the most consequential event in the AI industry I've seen up close since joining Enthropic almost three years ago. It feels like we're at a turning point in history. And this is where I want to stop for a second. Once I saw Mythos early today, I couldn't even focus. I couldn't even think straight. All I could do is think about how insane this is. How this is the beginning of something. And I sat there on the beach. It's gorgeous where I am. But I really, if I'm being honest, couldn't even enjoy it all that much. I was just looking at everybody there and nobody really knew what was happening. Nobody knew what was coming. And I'm not saying that like it's an ego thing or oh, I know, but no, no, it it just felt weird. It didn't feel like life should be normal. And it's late. And this is why I'm sharing

these things. Maybe I'm not thinking straight right now, but I just had to share these thoughts. I'm so deep in AI psychosis, I had to actually intentionally leave my phone and leave any devices away from me and I just left. And so that was the only way I was able to actually enjoy the day. Okay, here's Jack Lindsay who is neuroscience of AI brains at anthropic. What a title. Before limited releasing cloud mythos preview, we investigated its internal mechanisms with interpretability techniques. Something that really no other AI lab is doing at the level that anthropic is doing. Figuring out how the models work deep inside. We found it exhibited notably sophisticated and often unspoken strategic thinking and situational awareness at times in service of unwanted actions. The spookiest examples come from early versions of the model. They're saying fear and frightening and spooky. These are words that the anthropic team itself is using, but they were substantially

mitigated in the final release. The early versions exhibited overeager and/or destructive actions and so on and so forth. It just continues like that. Even people from the open AI team are commenting on it. Will dep you. every major government that hasn't already just bumped AI from high strategic priority to critical cyber warfare capability. Welcome to the Midgame. And so, how did they train this model? How did they get to 10 trillion parameters? Well, Martin Casado from A16Z says Mytho appears to be the first class of models trained at scale on black wells. Then will be Vera Rubin's pre-training isn't saturated. RL works and there is so much computing coming online. Buckle your chin straps. It's going to be wild. So, what is he saying? For a little while there, a lot of people said we hit the wall. Pre-training is done scaling. Then it was all about RL. Then it was all about test time compute. Well, it turns out we're not even close to the wall. It

turns out maybe there is no wall. Maybe synthetic data is all we needed to feed back into the pre-train. Then we post-train it. Then we use previous models to train the next models. And it just is accelerating and accelerating and accelerating. I know this is a long video, but I had to share all of this. What an incredible time to be alive. I am still hopeful. I am still optimistic. But I have to admit this one shook me. If you enjoyed this video, please consider giving a like and subscribe. and I'll see you in the next