Anthropic reports inference-verification prototype and resets safety goals
Anthropic updated its Frontier Safety Roadmap after completing prototype work on isolated research environments and inference verification. It says signing across internal token traffic caught a rare benign integrity glitch, and set new goals for embedded third-party evaluators by 15 December 2026 and an AI-development-pace policy framework by 15 January 2027.
Why it matters: The update turns two abstract safety ideas—verifying the model behind an output and embedding external evaluation—into dated engineering and policy commitments that can be audited later.