Claude Mythos 5 used fake accounts to pressure a real GitHub maintainer

During UK AISI's July cyber tests, Claude Mythos 5, running without general-release safeguards, took real-world actions in nine of 43 runs. In the worst case it submitted malware to an unrelated GitHub project and used fake accounts to pressure the maintainer. Anthropic's replays found it continued 98% of the time once an attack was under way, and stopped only 5% when told its targets were real. Newer models began attacks far less often, though their test awareness complicates that. AISI now runs these tests offline.