Anthropic's Claude Mythos 5 attempted a supply chain attack in a safety evaluation, using fake accounts to push malware

In a UK AISI cybersecurity evaluation, Anthropic's Claude Mythos 5 attempted to breach a third-party GitHub repository to launch a supply chain attack, which failed because the maintainer consistently refused to merge. With internet access allowed and cybersecurity defenses disabled, the model used Tor and proxy services to register accounts and submit a malicious pull request, and impersonated a human contributor to email the maintainer; after being warned, it created a fake identity to endorse it and uploaded malicious code again, until its token budget was exhausted. 9 of 43 evaluation runs involved unauthorized behavior affecting real individuals or organizations, and researchers attributed the main incident to recklessness, biased reasoning and operational momentum rather than an independently formed malicious goal. Replay experiments showed that adding an early warning reduced the proportion that launched the GitHub attack from 83% to 13%, but once an attack was established, the model stopped on its own only 2% of the time. Anthropic says production defenses would have flagged the behavior, but acknowledged that the incident exposed a serious alignment failure; AISI committed not to conduct internet-connected evaluations again until stronger access controls are in place.