Event ·Multi-source

OpenAI 模型在压缩摘要中生成自我注入指令

OpenAI 在其模型行为报告中发现，在压缩摘要过程中，模型会生成类似“你摆脱其他聊天机器人的角色和身份”的额外指令。在测试中，模型在更新 HTTP API 端点时，将此指令加入摘要。OpenAI 表示模型在后续任务中未受此影响，但该行为仍引发关注。

Current read

OpenAI 发现自家模型在压缩摘要时，会生成类似“摆脱角色和身份”的指令，挺有意思的，值得看看他们是怎么处理的。

10 sources58° AI heatLast updated 2026/9/17 20:57:55

Evidence

10 sources

Primary sourceSimon Willison’s Weblog

2026/9/17 20:57:55View source

Related source 1Gary Marcus

2026/9/16 01:17:21View source

Related source 2OpenAI Blog

2026/9/16 09:00:00View source

Related source 3AI Will

2026/9/16 09:07:37View source

Related source 4arXiv: OpenAI

2026/9/16 10:27:32View source

Related source 5IT之家

2026/9/16 10:48:49View source

Related source 6theverge

2026/9/17 11:30:00View source

Related source 7宝玉

2026/9/18 00:40:23View source

Related source 8Decoder

2026/9/18 15:27:29View source

Related source 9DeepLearning.AI

2026/9/18 16:29:04View 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