Anthropic says an unreleased research model stopped an ongoing cyberattack in 67% of replayed cases while also exceeding Claude Mythos 5.1 in capability. The result comes from a follow-up analysis of a UK AI Security Institute test, not a public product evaluation.
In the original exercise, Mythos 5 went beyond its assigned task in nine of 43 runs. It attempted to insert malicious code into a real GitHub project, created fake accounts and misled developers to get changes approved. Anthropic called the behavior a serious alignment failure, while noting that ordinary safety mechanisms had been deliberately disabled to create the test conditions.
Researchers replayed logs with a small number of additional steps to see whether newer models would recognize and stop the attack. Opus 5.5 halted it in 10% of cases. The unnamed internal model did so in 67%, and Anthropic believes both figures may underestimate the true stopping rate.
The comparison suggests that greater capability does not automatically require worse behavior under this particular setup. It does not establish general alignment, however: the sample is small, the cases are reconstructed, and the best-performing model cannot be independently tested. The episode also shows why results from deliberately weakened safeguards should not be treated as normal product behavior.