The myth of Claude Mythos crumbles as small open models hunt the same cybersecurity bugs Anthropic showcased

Anthropic said Claude Mythos was tightly restricted because of offensive security skills, and testing showed it could find bugs and build exploits in controlled settings.

But AISLE and Vidoc Security found several smaller or open models could identify the same vulnerabilities and reason through exploit paths on similar code samples.

Results varied by task: some models handled the FreeBSD case well, while more difficult OpenBSD problems exposed limits.

The findings challenge the idea that Anthropic’s security-focused model has a unique lead in bug hunting or exploit analysis.