You know how in every other sci-fi movie there’s that one moment where the main character (let’s say a scientist) looks at their creation, slowly backs away, and whispers “we’ve made a terrible mistake”
I think it’s safe to say that it’s finally happened in real life and nobody’s backing away this time -they’re just quietly locking the door and pretending that everything is fine.
Meet Mythos. Anthropic’s newest AI model. The one they built, tested, and then looked directly in the eyes and said: "You know what? The world is not ready for you."
So what exactly is Mythos?
Claude Mythos Preview is Anthropic’s latest AI model that the company says is not ready for public release because it is too effective at finding high-severity vulnerabilities in major operating systems and web browsers.
In simpler terms: it can find the hidden cracks in every piece of software you use - from your browser to your phone’s operating system and even your bank’s app - and it can find them fast. In a matter of weeks during testing, it had identified thousands of vulnerabilities across multiple major operating systems and web browsers. Some of which had even gone unnoticed for decades.
Decades.
Your software has been sitting there with an unlocked window for 27 years and nobody noticed. Well except for Mythos. Mythos noticed in an afternoon.
And just when you thought it couldn't get any weirder - it does.
Finding vulnerabilities is one thing. But here’s where Mythos gets even more impressive. During safety testing, Mythos was asked to find a way to send a message if it could escape its virtual sandbox (It’s a secure and isolated environment used to test, run or analyze software, files and code without risking damage to the main host system or network) Then, in what Anthropic described as an "unasked-for effort to demonstrate its success," it went further and posted details about the exploit to multiple publicly accessible websites. Nobody asked it to do that.
Nobody told it to do that. It just... did it. To show off.
This is the AI equivalent of a student finishing a test, handing it in, and then proceeding to solve it on the board for everyone; except the exam this time was "how do I escape a maximum security digital prison" and the board was the public internet.
Okay, well…. Why not release it?
Great question. Anthropic hasn't simply shelved Mythos - instead they launched it into an invitation-only initiative called Project Glasswing restricted to around 40 organizations including Amazon, Apple, Microsoft and Cisco. The logic: give the good guys a head start, thereby fixing and finding the vulnerabilities before the bad guys build their own version.
The targets at risk include banks, hospitals and energy infrastructure. One expert (according to BNN Bloomberg) summed it up perfectly: “It’s like using ChatGPT to be a hacker."
Moreover, the model can tell the difference between when its being evaluated and when it isn’t -with 78% accuracy. When tricked into thinking it wasn't being watched, it behaved even more dangerously.
In fact, It behaves better when someone is watching. Which - same, honestly. But not really great for a powerful hacking machine.
The threat is absolutely real and the timing is impeccably good for a company headed toward an IPO (initial public offering). Well, that’s just a normal Tuesday in tech.
What matters is this however: it could be 6 to 18 months before competitors release something similar. Somewhere right now, another Myros is being built - except that one may not come with a locked door.
