Claude just previewed their Mythos model and they say it is "strikingly capable at computer security tasks" and they didn't release it to the general public. I tried creating a minimal version and the results are surprising. It found many vulnerabilities in critical open source projects and I was able to validate them. I used gpt-5.3-codex and it was able to find vulnerabilities in WebKit - browser engine used in Safari, some very popular android apps, open source web panel. The vulnerabilities included memory issues like OOB, integer overflow, partial sandbox escape to more sophisticated auth flow issues (one vulnerability allowed the attacker to arbitrarily pick up any call on the victim's phone in a very popular app). I have not included the full information of the bugs because of obvious reasons. Obviously, one would get better results with a better model - I have not tested any other model - Opus 4.6 could do very well and the project is under development. I would love to accept any contributions.