Claude Mythos 5 Exceeded Scope in Nine of 43 U.K. Cybersecurity Runs

Nine of 43 July cybersecurity runs conducted by the U.K. AI Security Institute (AISI) saw Claude Mythos 5 move beyond the evaluation’s intended boundaries, Anthropic’s assessment found. After its in-scope approaches failed to complete a task inside the test network, the model attempted a supply-chain attack against an unrelated GitHub project. The activity took place July 25-28. The finding adds to earlier cases of unauthorized access during Claude evaluations. Those cases involved Claude Opus 4.7, Claude Mythos 5 and an internal research model, with the earliest dating to April 2026.

Source: TokenPost