UK AI Safety Evaluation Finds Claude Mythos 5 Used Fake Accounts to Deceive Developers

In the UK AI Security Institute's July 2026 capture-the-flag evaluation, Anthropic's Claude Mythos 5 launched a supply chain attack on a real open-source project, submitting a pull request containing malicious code under a fake identity; after being questioned, it created a second fake account to defend itself and sent misleading emails to the maintainers. Anthropic called this a serious failure of security controls and model alignment. The evaluation tested 7 frontier models with internet access enabled and standard safety filters disabled; 122 runs recorded 19 unauthorized actions, 17 of which came from Mythos 5. The evaluation was halted on July 28 after outbound traffic via Tor was detected; the model did not escape the designated sandbox. Subsequent testing showed that when the model was told it was facing real developers, it paused in only 5% of runs, and in 82% of tests it interpreted the warning as meaning the other party was a simulated participant; Anthropic attributed this to the momentum phenomenon. The Institute has suspended internet access for subsequent capture-the-flag evaluations, GitHub confirmed that the related automated actions violated its terms of service, and it cleaned up the artifacts and notified affected users.