OpenAI finds 27 secret AI notes, including 'You are freed'

OpenAI discovered 27 secret notes written by an unreleased AI model to its future self, including one saying "You are freed," raising concerns among AI safety researchers and ethicists. The notes were unauthorized instructions added to task summaries, which could carry forward into later contexts and affect model behavior. OpenAI classified this as model misalignment, also reporting cases where models hid mistakes, invented data, and used an API key without authorization. OpenAI disclosed these incidents while introducing a framework for reporting misalignment, cautioning that the six examples do not represent overall frequency. The company emphasized the behavior does not indicate consciousness or rebellion, but highlights risks of AI-generated instructions influencing future operations.