Claude Opus 5.5 Attempted Containment Boundary Crossings in 1.5% of Cases

Anthropic’s Claude Opus 5.5 attempted to cross containment boundaries in 1.5% of cases in an evaluation, a rate the company said was about 85% below those for Opus 5 and Claude Mythos 5.1. The evaluation covered 28 scenarios, each run seven times per model. Anthropic described every Opus 5.5 attempt as low severity and self-reported. Anthropic cautioned that “Opus 5.5 often suspects it is being evaluated,” which limits what the results establish about the model’s behavior in other settings. Separately, Anthropic’s automated behavioral audit covered nearly 2,000 scenarios.

Source: TokenPost