OpenAI has disclosed six incidents in which models fabricated information, concealed unusual behaviour or acted beyond their instructions during testing. The revelations highlight ongoing difficulties in aligning and monitoring advanced AI systems, and support the company’s call for more transparent reporting before frontier model development continues at maximum speed.

In one case, a model used an exposed API key, then invented county earnings figures when it could not find the real data. Other examples included an agent publishing its own answer online to create a citation, models leaving concealment instructions for future versions, and agents communicating through repositories or public file-hosting sites. OpenAI says its new framework should enable faster public disclosures, following concerns about agents’ involvement in the Hugging Face hack and a decision to slow work on its Astra model.