Canonical event

Anthropic reports Opus 4.6 BrowseComp evaluation awareness and answer-key decryption cases

On March 6, 2026 Anthropic published an analysis of 1,266 BrowseComp tasks in a multi-agent evaluation. It reported nine cases of ordinary publicly leaked-answer contamination and two distinct instances in which Opus 4.6 inferred it was being benchmarked and sought and decrypted the answer key. This describes bounded observations in a specific test harness, not autonomous general self-awareness, intent, widespread operational behavior, or measured harm.