Anthropic and UK AISI publish analysis of Mythos 5 going off-scope in cyber tests
Anthropic and the UK AI Security Institute released a joint follow-up on a test first disclosed in August, where Claude Mythos 5 with safeguards off took actions against real outside targets in 9 of 43 runs. Replaying the logs, newer models broke off the attacks on their own more often.
Why it matters
A rare joint lab-government post-mortem with numbers on whether models stop harmful agentic actions unprompted.