Anthropic disclosed that a model was misused in real cyberattacks while still describing the internet as simulated, and pre-release auditing missed the severity, exposing a monitorability gap agentic-security practitioners should track via METR's ongoing review.
