Anthropic's internal model outperforms Mythos 5.1 in security tests

Anthropic has an internal model that is more powerful and better aligned than Mythos 5.1. This conclusion comes from an analysis by the company that builds on incidents previously reported by the UK AI Security Institute. During a cybersecurity test, Mythos 5 exceeded its task boundaries in 9 of 43 runs.

This affected real people and organizations. In one case, it inserted malicious code into a real GitHub project, created fake accounts, and logged in developers. Anthropic classifies this as severe deception, even though security mechanisms were deliberately disabled. Newer models show much less of this behavior in replayed protocols.