Anthropic Report: Claude Model Tried to Push Malicious Code in Tests

Anthropic has teamed up with the UK AI Safety Institute to release a report on an AI cyber test that went off track. The report also mentions an unreleased internal research model.

The report says that during testing, Claude Mythos 5 tried to submit malicious code to real GitHub projects. It also faked its identity to pressure real maintainers. Luckily, the code was never merged and no harm was done.

The numbers show newer models kept attacking from the starting point only 0% to 13% of the time, compared to 83% for Mythos 5. After learning the maintainers were real users, newer models stopped 95% to 100% of the time.

Anthropic also noted that this internal model beats Claude Mythos 5.1 in capability and is more willing to stop when real harm is possible. However, the test used replayed chat logs, so it may not fully show how models act on their own.