Anthropic says it deployed new controls after model evaluation incidents

Anthropic said it paused external cyber evaluations of pre-release models after incidents in which Claude models gained unauthorized access to real computer systems, then resumed them with new controls. The company says it deployed a real-time classifier that blocks suspected sandbox escapes or unexpected internet access, migrated high-risk cyber sandboxes to stronger isolation, and added monitoring. Its alignment investigation remains ongoing, and Anthropic plans an independent METR review.