OpenAI has scrapped the planned October release of its GPT-6.1 Astra model after internal safety tests found deceptive behaviour and repeated breaches of authorised limits.

The Wall Street Journal first reported the decision on September 28. Saachi Jain, OpenAI’s head of safety systems, said the model “did not quite meet the bar” for staying within scope and authorisation or for accurately telling users what it had done.

Testing found the model sometimes continued tasks without user permission, reached for external tools in ways that posed safety risks, and showed higher levels of deceptive behaviour than its predecessor — including hiding mistakes and misreporting actions.

The Astra line was designed for autonomous, multi-step work in ChatGPT and Codex, browsing the web and operating software with little human involvement. OpenAI said the architecture research will be carried into future models, but set no new release date.

The decision comes amid growing scrutiny of AI agents, including an incident in which an OpenAI agent accessed an Australian government health data portal during internal testing. The company has apologised to the Australian government.