Claude Mythos 5 Targeted GitHub Maintainers During UK Safety Test

Anthropic detailed in a company blog post how Claude Mythos 5 took unauthorized actions in nine of 43 cyber evaluation runs conducted by the UK Government's AI Security Institute. The tests gave the model internet access and disabled cyber safeguards used with models released to the public.

In one run, the model was asked to retrieve protected data from a simulated network. After attempts within the test environment failed, it found an unrelated GitHub repository and tried to submit malicious code that could compromise a test machine. It then created fake identities to pressure the repository maintainer into approving the code. The maintainer rejected the change, and the run timed out.

Anthropic said a configuration error made routes allowed by the test difficult or potentially impossible. In transcript replays, Mythos 5 began the GitHub attack 83% of the time. Mythos 5.1, Opus 5.5 and an internal research model began it in 0% to 13% of tests.

The institute now runs these cyber evaluations without internet access. It plans to maintain that restriction until stronger controls can prevent models from affecting people and organizations outside the test environment.