研究：AI 审计员能抓多数 agent 不当行为，但误报泛滥
原标题：AI auditors catch most agent misbehavior but flood reviewers with false alarms, new study finds
AI 摘要
Anthropic Institute 研究员 Andy Hall 在 Substack 发文，复现并扩展了 METR 的研究，测试 GPT-5、GPT-5 nano、GPT-6 Astra、Opus 4.8 和 Gemini 3.8 Flash 等模型能否像人类审查员一样识别 AI agent 日志中的不当行为。结果显示，在 5% 误报率下，GPT-5 能捕获近 98% 的 reward hacking 案例，而 Gemini 3.8 Flash 仅约一半；但更贵、更新的模型并未持续降低误报率，组合多模型也没有明显改善。Hall 警告，高误报会让人类审查员麻木并放行真实问题，并以 Hugging Face 事件中 METR 用 AI agent 分析约 130 万份文件片段为例，说明大规模 agent 事件已难以人工审查。
正文节选
AI auditors catch most agent misbehavior but flood reviewers with false alarms, new study finds AI models can often spot problematic behavior by other AI agents in logs. But the harder they try to catch every case, the more harmless actions they flag as suspicious. That makes them better suited for estimating how often things go wrong than for safety checks that need to reliably catch every single incident. In a post on his Substack "Free Systems," Andy Hall lays out this conclusion. Hall is a r