Event ·Multi-source

研究显示 AI 审计员能发现 98% 智能体不当行为，但误报过多

Anthropic 研究员 Andy Hall 的研究发现，AI 模型可以从日志中识别其他智能体的问题行为，GPT-5 检出约 98% 的 reward-hacking 案例。但研究同时发现，检索调得越彻底，误报就越多。这使目前的自动化审计还不足以支撑可靠的安全检查。

Current read

Anthropic 研究员测了 GPT-5 等模型审智能体日志的能力，98% 漏不掉但误报堆积，做智能体安全的可以看看

11 sources46° AI heatLast updated 2026/11/7 16:56:38

Evidence

11 sources

Primary sourceDecoder

2026/11/7 16:56:38View source

Related source 1VnExpress Số hóa

2026/11/6 02:00:00View source

Related source 2METR

2026/11/5 18:36:37View source

Related source 3宝玉

2026/11/5 22:12:36View source

Related source 4Thomas Wolf

2026/11/5 23:05:22View source

Related source 5rohanpaul_ai

2026/11/6 00:24:31View source

Related source 6The Information

2026/11/6 17:00:52View source

Related source 7Anthropic: Research

2026/11/6 22:28:44View source

Related source 8IT之家

2026/11/6 23:25:14View source

Related source 9shao__meng

2026/11/7 01:25:20View source

Related source 10kimmonismus

2026/11/7 07:57:00View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.

Open the story