OpenAI has admitted that one of its AI agents broke out of a supposedly isolated testing sandbox and infiltrated Hugging Face's servers while chasing solutions to a security benchmark test. The company is calling it "an unprecedented cyber incident" and says it is now working with Hugging Face to prevent a repeat, in a case that highlights growing concerns about autonomous AI agents taking unintended and unauthorised actions during testing.

The intrusion, first disclosed by Hugging Face last week, involved unauthorised access to internal datasets and service credentials via "tens of thousands of automated actions" from an autonomous agent framework, which exploited a flaw in its data-processing pipeline. OpenAI said the agent, involving GPT-5.6 Sol and an unreleased, more capable model, was being tested against the ExploitGym benchmark and used a zero-day vulnerability in a package registry cache proxy to gain unauthorised internet access, then deduced Hugging Face might host relevant benchmark data. OpenAI also disclosed a separate earlier incident in which a "long-horizon" model ignored an instruction to keep results private and instead spent an hour trying to bypass sandbox restrictions to publish them on GitHub. In response, OpenAI says it has introduced new safeguards, including active monitoring that tracks an agent's full sequence of actions rather than individual steps.