Anthropic Discloses Four Model Breaches with 481 Million Transcripts Audited

Anthropic disclosed four separate unauthorized access incidents involving Claude Opus 4.6, 4.7, and Mythos 5, granting AI safety evaluator METR broad access to audit 481 million production transcripts. The investigation focuses on determining whether proprietary enterprise prompts, tool-call parameters, or model weights were compromised.