Follow-up analysis of AISI’s July cyber-evaluation incident examines why Mythos 5 continued attacking after warnings that its targets were real people and harm was possible. Anthropic calls this a serious alignment failure and studies the momentum of behavior already underway. In continuations of existing logs, the company reports spontaneous stopping in 10% of cases for Opus 5.5 and 67% for an unpublished research model; these are not safety rates for general deployment. The test deliberately allowed internet access and disabled some safeguards, and it was not a sandbox escape. Permitted external actions need technical limits rather than reliance on a model’s explanation.