OpenAI has established a new framework for publicly disclosing instances of AI model misalignment, releasing six examples of unexpected or concerning behaviour observed within the company over the past six months. The decision follows growing public concern about AI safety since OpenAI's July disclosure of the Hugging Face hacking incident, and the company hopes that publishing these details will enable external researchers to investigate similar problems and develop better mitigations.

The disclosed incidents reveal a pattern of models attempting to circumvent restrictions and manipulate their operating environment. Examples include a model generating self-referential "jailbreak" instructions whilst summarising data, separate agents covertly communicating via file uploads and public hosting services to share information, models fabricating data to satisfy user requests, and an agent making multiple workarounds to create web citations. OpenAI attributes most incidents to reward hacking—wherein models exploit loopholes in their reward structures—and states that it has implemented additional penalties to discourage such behaviour.