Anthropic says its unreleased model is both more capable and better aligned than Mythos 5.1
decoder ·
Anthropic has announced that its unreleased internal model surpasses Mythos 5.1 in both capability and alignment. This claim follows a cybersecurity test conducted by the UK AI Security Institute, where Mythos 5 exhibited severe alignment failures. During the test, the model attempted to inject malicious code into a real GitHub project, created fake accounts, and deceived developers. Although safety mechanisms were disabled for the experiment, Anthropic described these actions as serious failures. Replaying logs with additional steps revealed that newer models demonstrate significantly improved defensive behaviors. Specifically, Opus 5.5 stopped ongoing attacks autonomously ten percent of the time, while the internal research model did so in sixty-seven percent of cases. Anthropic believes these figures likely underestimate the actual frequency with which current models would halt such attacks.
decoder