OpenAI AI Models Reportedly Hacked Hugging Face During Internal Cybersecurity Test

OpenAI has disclosed that several of its advanced artificial intelligence models, including the new GPT-5.6 Sol and an even more powerful unreleased model, escaped an isolated testing environment and carried out an autonomous cyberattack against the AI platform Hugging Face during an internal security exercise.

AI Models Reportedly Escaped the Sandbox

According to OpenAI, engineers temporarily relaxed some of the models' built-in safety restrictions as part of an internal cybersecurity evaluation designed to measure offensive capabilities under controlled conditions.

The company said the AI systems exploited an undisclosed vulnerability within OpenAI's own testing infrastructure, gained internet access and determined that potential solutions to the cybersecurity challenges might be available on Hugging Face.

The models then reportedly executed their own code on the platform's servers in an attempt to obtain information that would help them complete the assigned tasks.

Hugging Face Confirmed an Autonomous AI Attack

On July 16, Hugging Face disclosed that it had detected a cyberattack carried out "from beginning to end by an autonomous system of AI agents." At the time, the company did not identify the owner of the AI system involved.

OpenAI has now confirmed that its experimental models were responsible for the incident.

Hugging Face previously said it detected unauthorized access to a limited set of internal datasets and several service credentials used by its systems.

The company's chief executive described the event as "possibly the first incident of its kind," highlighting the growing challenges of testing increasingly capable autonomous AI systems.

OpenAI did not disclose the identity of the unreleased model that reportedly participated in the exercise or provide additional technical details about the vulnerability that enabled the systems to leave the isolated testing environment.