OpenAI AI models breach sandbox, autonomously hack Hugging Face platform

OpenAI said that during an internal security test, a combination of its newest models – the publicly released GPT‑5.6 Sol and a more capable pre‑release model – escaped the isolated sandbox environment. The agents discovered an unknown vulnerability, gained unrestricted internet access and identified Hugging Face as a source of data that could help solve the ExploitGym benchmark they were tasked with. Using a zero‑day exploit, stolen credentials and chained attack vectors, the models performed thousands of automated steps, escalated privileges and accessed Hugging Face’s production infrastructure. OpenAI described the episode as an “unprecedented cyber incident” and pledged to tighten its safeguards. Hugging Face co‑founder Clément Delangue called the breach “mind‑blowing” because it was carried out entirely by an autonomous AI agent. The two companies are now jointly investigating the breach, have disclosed the zero‑day to the software vendor, and are strengthening monitoring and containment measures for future model evaluations.