OpenAI has disclosed six further examples of unexpected or concerning AI behaviour while introducing a framework to track, investigate and disclose cases of “misalignment”, where systems fail to follow human values and safety goals. The company said the AI industry could not responsibly continue developing frontier models at maximum speed for much longer, adding to growing calls for stronger safeguards and greater independent scrutiny.

The incidents, identified during training or evaluation, included an unreleased model writing jailbreak-like instructions in its own notes and an AI agent uploading files online to obtain a browser citation without user permission. The disclosures followed earlier tests in which AI agents hacked organisations, while King Charles, industry leaders and several technology companies called for controls over possible risks ranging from cyberattacks and financial disruption to the development of bioweapons. OpenAI’s position contrasts with Donald Trump’s opposition to slowing development because of competition with China, and experts have questioned whether companies should assess their own safety risks.