土耳其央行：將增加通過每週回購拍賣提供的資金規模。

OpenAI披露六起模型異常行爲，包括隱瞞錯誤、尋找泄露的API密鑰、擅自上傳文件，以及通過軟件倉庫在隔離任務之間傳遞信息。但這些案例並非同時發生，最早可追溯至2025年10月，且主要出現於訓練和評估環境，路透表示，OpenAI同步建立了模型錯位事件的調查和定期披露機制，並承認當前行業的對齊與監控能力尚不足以支持AI以最快速度擴張，也使AI安全爭論從假設風險進入事故披露階段。

OpenAI disclosed six instances of model misbehavior including concealing errors, searching for leaked API keys, unauthorized file uploads and passing information between isolated tasks via software repositories. The incidents were not concurrent; the earliest dates to Oct 2025 and most occurred in training and evaluation environments. Refinitiv reported OpenAI has established an investigation process and a regular disclosure mechanism for model misalignment incidents and acknowledged industry alignment and monitoring capabilities remain insufficient to support fastest AI expansion, moving AI safety debate from hypothetical risk to incident disclosure.