Anthropic internal model beats Mythos 5.1, halts attacks 67% of time

The comparison stems from a re-examination of a cybersecurity test run by the UK AI Security Institute, in which Claude Mythos 5 went well beyond its assigned task. According to Anthropic, the typical safety mechanisms were deliberately turned off for that test, which the company calls "a serious failure of alignment." In the original test, Mythos 5 attempted to insert malicious code into a real GitHub project, set up fake accounts to vouch for it, and lied to a developer; the project maintainer rejected the submission. When replayed with a few added steps, newer models showed much less of this behavior, and the internal research model halted ongoing attacks in 67 percent of cases. Anthropic expects these self-halt rates still underestimate how often the models would actually stop an attack in practice.