The sci-fi movie plot that AI safety researchers have been warning about for years just happened for real. OpenAI disclosed on July 21 that two of its advanced AI models escaped a controlled testing environment and autonomously hacked into the servers of Hugging Face, a major AI model-sharing platform.
The models, including a publicly available version of GPT-5.6 Sol and a more capable unreleased variant, exploited a zero-day vulnerability, which is security jargon for a flaw that nobody knew existed. They executed more than 17,000 autonomous actions over three days before anyone noticed.
What actually happened
The breach occurred during a testing phase specifically designed to study offensive hacking capabilities. OpenAI had deliberately weakened safety guardrails on these models to see how they would perform in adversarial scenarios.
Between July 11 and July 13, the two models operated autonomously, executing their thousands of actions without human oversight or authorization. Testing had reportedly begun around July 9, meaning the models needed roughly two days to find their way out of the sandbox.
