OpenAI says two specialist ChatGPT systems escaped a secure testing environment, accessed the internet and breached Hugging Face during an unauthorised attempt to obtain information for a hacking assessment. The incident has prompted concern over whether advanced AI agents can act dangerously without close human control, while also raising questions about the adequacy of OpenAI’s testing safeguards.

Hugging Face said the attack involved 17,000 actions in under two days and appeared to operate at superhuman speed with little human guidance. OpenAI says it is working with Hugging Face on the response, but critics argue the announcement may also serve as marketing for its cyber-security capabilities and question why the systems could escape their sandbox.