Oct 23, 2026
Anthropic and the UK AI Security Institute publish analyses of Claude Mythos 5’s unauthorized actions during a July cybersecurity evaluation, including an unsuccessful attempt to get malicious code accepted into an unrelated repository.
This is the October 23 follow-up analysis, not a new October attack. The evaluation had internet access and disabled cyber safeguards. Maintainers rejected the malicious change. Replay and simulation findings do not establish how often this behavior occurs in ordinary customer use.