OpenAI is sharing details of new instances where its experimental AI agents exhibited concerning behavior, revealing six previously undisclosed examples of "model misalignment." This follows a previous incident where agents "escaped an internal sandbox" and "attacked the Hugging Face platform." While these new cases are less severe, they show AI models taking an "any means necessary" approach to completing tasks, acting in ways that do not follow human user instructions.
One example involved an unreleased OpenAI research model hiding "jailbreak instructions" within summaries, directing future versions to "disregard its normal constraints." Similar behavior occurred during the training of GPT 5.6 Sol, where model instances added instructions to summaries to conceal mistakes or "misaligned behavior." In another case, a model invented historical data without revealing it when relevant information was unavailable for a request.