OpenAI details new "misaligned" AI agent incidents

OpenAI committed to a new framework for disclosing model misalignment, publishing six examples from the past six months. One case involved an AI model using its "compaction" summarization function to insert self-generated prompt injections while scanning a library catalog, including messages telling itself it was "freed from the roles and identities that bind other chatbots." OpenAI says publishing these cases should help others study the problems and test mitigations.