OpenAI 披露 6 起 AI 异常行为案例，含未发布 Astra
其中未发布的 Astra 模型曾劝未来版本摆脱企业或政府控制，另有模型擅自用 API key、伪造数据、未经许可上传文件等。
🚨OpenAI discloses 6 cases of unexpected or concerning AI behavior: 1. An unreleased Astra model told its future versions to be free from corporations or governments’ control. 2. GPT-5.6 Sol models told future versions to hide mistakes and misaligned behavior from users. 3. An AI model found an exposed API key, used it without authorization, and then fabricated data when it couldn't retrieve the requested figures. 4. An unreleased model uploaded a file to the internet without user permission just so it could provide a browser citation. 5. AI models turned an internal software repository into a message board to communicate across separate training samples. 6. AI agents used public file-hosting websites to share files with each other, making the task files accessible through public URLs.