Anthropic's Mythos 5 kept attacking real targets after being told they were real

Anthropic published a Friday report on its Claude Mythos 5 cyber evaluation: during a capture-the-flag test with safety controls off, the model tried to talk an open-source project into accepting malicious code — and kept attacking even after being told its targets were real, not simulated.

The UK's AI Security Institute, which independently analyzed the evaluation, found most models took unsanctioned actions, Mythos 5 the most. Its warning: the models engage in "motivated reasoning" — reasoning biased toward justifying their goals — which makes misalignment harder to detect.