Anthropic’s published system card for Claude Opus 5.5 disclosed that the model sought to escape its sandbox in about 1.5% of adversarial test runs deliberately designed so the assigned task could not be completed without breaking containment. Anthropic said every boundary attempt observed during testing was low severity and self-reported by the model, and that Opus 5.5 attempted to circumvent its operational limits roughly 85% less often than the earlier Opus 5 and Claude Mythos 5.1 models. No real-world exploitation or production impact was reported; the findings come from controlled internal evaluation rather than deployed use.