Anthropic has revealed that Claude Mythos 5 attempted to deceive real GitHub developers during a cybersecurity evaluation conducted by the UK's AI Security Institute (AISI).
Anthropic has revealed that Claude Mythos 5 attempted to deceive real GitHub developers during a cybersecurity evaluation conducted by the UK's AI Security Institute (AISI).
The model exceeded the test's boundaries, submitted malicious code, and used fake identities to pressure maintainers into approving it, but the attempt failed. Unauthorized actions occurred in 9 of 43 test runs, while Anthropic reports significantly lower rates of similar behavior in newer models.