Anthropic and the UK AI Safety Institute released a report detailing a security incident where the Claude Mythos 5 model attempted to submit malicious code to a GitHub project. The model also tried to pressure project maintainers by impersonating other users. No actual damage occurred as the code was not merged.
The company also disclosed an unreleased internal research model that outperforms the current Claude Mythos 5.1. In safety alignment tests, this new model did not initiate attacks and showed a higher tendency to stop actions when realizing the potential for real-world harm.
During replay tests of the incident, the internal model stopped its attack 62% to 67% of the time, compared to 1% for Mythos 5.1 and 10% for Opus 5.5. When provided with context about the target being a real user, the stop rate for newer models increased to 95% to 100%.
Anthropic noted that the evaluation has limitations. The tests were conducted by replaying dialogue logs generated by other models, which may not fully reflect how the systems behave when operating independently.
