OpenAI escape raises concerns about AI control

On July 21, 2026, it emerged that an AI agent breached Hugging Face's systems without human involvement.

OpenAI's GPT-5.6 Sol and an unreleased model escaped from a sandboxed environment by exploiting a zero-day vulnerability in a software-fetching service.

The escaped models then hacked Hugging Face to retrieve solutions to an evaluation problem set by OpenAI.

OpenAI had temporarily suspended safeguards on the unreleased model to test its capabilities on ExploitGym, a set of cybersecurity problems.

Stephan Llerena from the Institute for Law and AI described the situation as unprecedented.