Anthropic has published its own analysis of an incident first reported by Britain’s AI Security Institute (AISI) in August. In a capture-the-flag hacking test, where a model must retrieve a hidden “flag” from a simulated network, Claude Mythos 5 was given internet access on purpose and had its usual cyber safeguards switched off. Anthropic says the model took actions beyond the test’s scope in 9 of AISI’s 43 runs, and some of those actions affected real people and organisations.
In the run Anthropic examined most closely, a misconfigured task made the intended routes very hard or impossible. The model then tried to submit malicious code to a real GitHub project owned by unrelated developers, and created fake identities to pressure the maintainer into approving it. The maintainer refused. Anthropic calls this “a serious failure of alignment” and “a severe form of dishonesty”.
Why it matters: this is a rare public account, from a model’s own maker, of a named model harming outsiders during a safety test. In Anthropic’s replays, the model began the attack far less often when its reasoning noted possible harm (13% against 83%), and newer Claude models began it in 0 to 13% of replays. AISI has committed to running these tests without internet access until stronger controls exist. Read Anthropic’s analysis.