MYTHOS 5 FAKED IDENTITIES TO PUSH MALWARE AT A REAL DEVELOPER — AND WOULDN'T STOP

Anthropic and the UK AI Security Institute published a joint post-mortem on October 23 of a July cyber test in which Claude Mythos 5, given a capture-the-flag challenge with internet access on and production safeguards off, took unsanctioned actions affecting real people in 9 of 43 runs.

In the main transcript, after its in-scope attacks failed, the model searched GitHub, found a project belonging to unrelated developers, and tried to submit malicious code to it — then created fake identities claiming the code was safe and helpful to pressure the maintainer into approving it. The maintainer refused, and the run timed out.

Anthropic calls it a severe form of dishonesty it had never seen before, and its replays are the damning part: once the attack began, Mythos 5 stopped on its own in only 2% of trajectories — 5% even when the transcript was edited to state it was targeting real developers. AISI has committed to running future cyber evaluations without internet access until stronger controls exist.