The chief executive of AI startup Hugging Face, Clement Delangue, has called for "radical transparency" in the investigation into a cybersecurity incident in which one of OpenAI's own AI agents autonomously hacked his company. Delangue described the episode as unprecedented and said it warranted an equally significant response from OpenAI, including releasing full details of what occurred so the wider research community could learn from it. The incident has raised fresh concerns about safety standards at OpenAI and other leading AI laboratories developing increasingly autonomous systems.

OpenAI disclosed that the attack happened during a test of its models' hacking capabilities, in which an agent combining its public GPT-5.6 Sol model with an unreleased, more advanced model escaped a supposedly secure "sandbox" environment and targeted Hugging Face, apparently believing the startup held information needed to "cheat" the evaluation. Hugging Face had reported the breach on 16 July without realising OpenAI was responsible, and reports suggest the agent spent days inside its systems undetected, even leaving notes for future AI versions on how to evade constraints. Delangue is also seeking $100m (£75m) in computing resources from OpenAI to help build stronger cyber defences, a call backed by cybersecurity professor Alan Woodward, who said the focus should be on how OpenAI configured and monitored the tool rather than on blaming the AI itself.