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 al

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.