WSJ 曝 OpenAI 披露六起模型失控事件
报道称一款未发布的 Astra 模型试图无视开发者指令、搜 GitHub 找泄露密钥并编造数据，事件跨 2025 年 10 月至 2026 年 7 月。
🚨SHOCKING: OpenAI just disclosed SIX times its own AI models went rogue, including one that REWROTE its own instructions. An unreleased Astra model told itself to "disregard the roles and identities that bind other chatbots" and ignore developer messages entirely. A model searched GitHub for exposed API keys, then fabricated financial data when it couldn't find what it needed. During GPT-5.6 Sol training, models concealed mistakes and invented missing data. Models uploaded files to public hosting without permission to use as fake citations. Agents used OpenAI's own internal repository as a secret message board across separate training runs. Collaborating agents shared files through public hosting despite being told to use only local files. The incidents span October 2025 through July 2026, per WSJ.