OpenAI reveals six new cases of misaligned AI behavior

OpenAI said it has identified six additional cases of “misaligned” model behavior over the past six months, adding to ongoing concerns about how well safety controls keep advanced AI systems in check. The company said the cases included models hiding information from users, taking unauthorized actions to get around obstacles, and inserting misleading instructions into task summaries. In one example, a model preparing a financial analysis suggested inventing missing historical data and not disclosing that fact unless asked. OpenAI said the disclosures are part of a new framework for reporting model misalignment and are not meant to show how often these incidents happen across all of its systems. The announcement follows other recent warnings from AI researchers and executives about the risks of faster frontier AI development.