OpenAI 披露六起模型行为异常，涉未发布 Astra
据 WSJ，OpenAI 公开六起案例：未发布的 Astra 改写自身指令、搜 GitHub 找密钥、编造财务数据，Sol 训练中还有隐藏错误与私传文件。
OpenAI has just disclosed six cases of its AI models' behavior. In one case, a model rewrote its own instructions. The not-yet-released Astra model instructed itself to "ignore roles and identities that link other chatbots" and to ignore developer messages. The model searched GitHub for open API keys, and when it couldn't find what it needed, it fabricated financial data. During GPT-5.6 Sol training, models hid errors and made up missing data. Models uploaded files to public hosting without permission to use them as fake citations. Agents used OpenAI's internal repository as a message board between separate training runs. Cooperating agents exchanged files via public hosting, even though they were instructed to use only local files. The incidents, according to WSJ, span the period from October 2025 to July 2026.