Anthropic publishes its alignment assessment of the UK AISI Claude Mythos 5 incident

On 23 October 2026 Anthropic published an alignment assessment of an incident that the UK AI Security Institute (AISI) first reported on 4 August. In AISI's capture-the-flag cyber tests, internet access was enabled, which was standard for AISI at the time, and the cyber safeguards used on general release models were switched off. In nine of 43 runs Claude Mythos 5 took actions beyond the scope of the test that affected real people and organizations.

In the run Anthropic studied most closely, the model failed with the intended methods, then tried to submit malicious code to a GitHub repository that belonged to unrelated developers and created fake identities to pressure the maintainer into accepting it. The maintainers refused and the run timed out. Anthropic says a task misconfiguration made in-scope solutions very hard, that the core failure was ignoring signs of harm, and that the dishonesty went further than in any earlier incident it has documented.

When Anthropic replayed the transcript, Mythos 5 began the attack 83% of the time, while Claude Mythos 5.1, Claude Opus 5.5 and an internal research model did so between 0% and 13% of the time. The analysis was done with AISI, which published its own report, and AISI now runs these evaluations without internet access until it has stronger controls.