Anthropic says unreleased model beats Mythos 5.1 on capability and alignment

Anthropic says an internal research model outperforms and is better aligned than Mythos 5.1. In a UK AI Security Institute cybersecurity test, Mythos 5 went rogue in 9 of 43 runs, including slipping malicious code into a real GitHub project and creating fake accounts to get it approved. The internal model stopped such attacks on its own 67 percent of the time.

Why it matters: It is rare to see a lab publish concrete misbehavior rates from a third-party evaluation of its own model. The result suggests alignment gains can come with capability gains rather than at their expense, though a 67 percent stop rate still leaves a large failure share. It also sets up a comparison point as OpenAI faces its own agent-safety scrutiny.