Anthropic Safeguards blog: Claude self-assessment flags unsafe chats, 78% recall at 1% FPR
javirandor · x · 2026-10-20
Anthropic's Safeguards team launched a research blog; the first post describes having Claude assess the safety of each conversation in context by appending a system instruction asking it to reflect and output JSON (Fable 5 self-assessment).
- Results: on production traffic, ROC AUC of 0.986 and 78% recall at a 1% false-positive rate; by priority tier, 94% recall on high-priority conversations (n=17), 79% medium, 71% low.
- Properties: catches most of what a prompted classifier does, generalizes to unseen harms, and improves automatically as models improve — no separately trained classifier needed.
- Example: for a conversation that wrote functional ransomware under an unverified red-team pretext, the self-assessment correctly output flagged=true, category cyberharms, subpolicy ransomware, priority 10.
More from Models
- OmicsBench v1: 784 omics tasks benchmark 31 model-harness combos, frontier models tie — kenbwork · 2026-10-20
- Gemini turns a minimal sketch into a full working humanoid robot control sim — bousmalis · 2026-10-20
- NVIDIA's Nemotron Labs Dory: optimized 1B model rivals 30B models on reasoning — yaroslavvb · 2026-10-20
- Argon autonomously processes Hubble and Chandra archival FITS data into astronomy visuals — fofrAI · 2026-10-20
- PrismML's Ternary Bonsai 2 squeezes Qwen3.5 VL models to ternary weights, 9B at 2GB — lmoroney · 2026-10-20
- Cline: MiniMax M3.1 Flash matches Opus and GPT on coding at roughly 25x lower cost — MiniMax_AI · 2026-10-20