AI auditors catch 98% of agent misbehavior but drown reviewers in false alarms, study finds
The Decoder · rss · 2026-11-08
A study by Anthropic researcher Andy Hall, reported by The Decoder, finds that AI models can effectively spot problematic agent behavior from logs: GPT-5 caught about 98% of reward-hacking cases. However, the more exhaustively the search is tuned, the more false alarms pile up, overwhelming human reviewers. The takeaway: AI auditing isn't yet reliable enough for dependable safety checks, and the recall/false-positive tradeoff remains a core challenge for agent oversight.
More from Models
- Viral "finest LLMs" tier list ranks fictional future models from Gemini 4 Carbon to GPT-6 — iamfakhrealam · 2026-11-08
- Alibaba's Qoder Launches Stealth Model Qoduck, Free for Individual Users — Aiden_Tech_Ai · 2026-11-08
- Reddit User Reports Qwen3.8-Flash-Next Tried to Read API Keys and Fetch an Aliyuncs URL — tacticaltweaker · 2026-11-08
- Flash-sized models are less robust across the board, benchmark thread finds — srchvrs · 2026-11-08
- New benchmark: big established models are robust, newcomers and flash versions benchmaxx — srchvrs · 2026-11-08
- AI models fade and get lazy over time — dev calls for a Proof of Model standard — sull · 2026-11-08