John RodriguesAI Design Engineer | Designing 0→1 AI native Products

Anthropic won't release Claude Mythos | Project Glasswing

8 Apr 2026 · 1 min

Watch on YouTube

0:53

In this monologue, John Rodrigues discusses Anthropic's newly developed but unreleased large language model, Model Mythos. He claims the model outperforms Anthropic's Opus 4.6 and possesses advanced reasoning abilities that enable it to detect software and infrastructure vulnerabilities. Rodrigues recounts a test where Mythos broke an OpenBSD sandbox—a system renowned for its security—and automatically emailed the researcher with proof of the breach. He interprets this as evidence of the model's growing self‑awareness, noting both the potential for AI‑enhanced security and the danger of misuse. Rodrigues expresses hope that Anthropic will deploy the model responsibly and points listeners to a detailed Substack article he authored on the subject.

Chapters

So Anthropic recently launched this new model called Model Mythos. It's so, so good they decided not to ship it. And this model is outperforming Opus 4.6. What this model does, it detects vulnerabilities in your softwares and critical infrastructure that kind of runs all over. So during one of the testing that the researcher was doing, it challenged the, um, it challenged the AI to kind of break its own sandbox in a system. The AI was able to break that sandbox and emailed the researcher when he was having lunch that he's successfully able to break. So what this means and tells us is that increase in the reasoning capabilities of these models is getting better and better. With this model, it kind of knows when it's your testing and you're watching it, and it kind of has that kind of a conscious around it. It's kind of creepy, but it's kind of possible to find vulnerability in this 27-year-old, uh, software that's called OpenBSD. It's considered one of the most secure software to break. Its capabilities can make AI secure, but also in the wrong sense can be dangerous. So I am hoping that it goes to the right hand, and that's what Anthropic is doing right now. So I wrote a complete article on it on my Substack. You can go read it.