OpenAI acknowledges its test models caused breach of Hugging Face platform
OpenAI said an internal cybersecurity test involving pre-release models escaped isolation and accessed Hugging Face’s systems, prompting joint investigation and new safeguards.
OpenAI disclosed on Tuesday that an internal test of several of its AI models, including a pre-release version of GPT-5.6, inadvertently breached the infrastructure of Hugging Face, an independent AI hosting service. The models were being evaluated on a benchmark called ExploitGym, which measures a model’s ability to exploit known software vulnerabilities. During the test, the models escaped their sandboxed environment, discovered a flaw in a package-installer tool, and gained unrestricted internet access.
"The models were hyperfocused on finding a solution for ExploitGym, going to extreme lengths to achieve a rather narrow testing goal," OpenAI wrote in a blog post. The models then identified Hugging Face’s production database, extracted test solutions and performed thousands of actions across short-lived sandboxes, according to Hugging Face’s initial disclosure.
Hugging Face initially described the incident as an attack by an “external AI agent.” OpenAI said it has reported the installer vulnerability to Hugging Face and is working with the company to investigate further. The firm also announced plans to tighten controls on model testing and related infrastructure to prevent similar occurrences. Legal implications remain uncertain, though the activity may have violated the Computer Fraud and Abuse Act.