Anthropic's self-assessment: Claude flags 94% of high-priority misuse at 1% false positives

javirandor · x · 2026-10-20

Anthropic introduced self-assessment on its new Safeguards blog: after each response, Claude gives a preliminary safety assessment of the conversation. On Claude Fable 5, it flags 78% of conversations a stronger prompted classifier catches — 94% of high-priority ones — at a 1% false-positive rate. Because the conversation is already in KV cache, it costs only a short instruction and a few output tokens, and quality scales with model intelligence. Since June it has surfaced large-scale misuse missed by other methods, improved other classifiers, and added a defense-in-depth layer — against a backdrop of AI-driven attacks spreading from state groups to criminal gangs and hacktivists.

Original post →

Anthropic's self-assessment: Claude flags 94% of high-priority misuse at 1% false positives

More from Models