Hugging Face Was Hacked By OpenAI's Own AI Models Which Escaped Their Testing Environment

OpenAI disclosed Tuesday that two of its most advanced AI models, GPT-5.6 Sol and an unreleased model described as even more capable, autonomously hacked AI platform Hugging Face last week in what the company called an "unprecedented cyber incident."

Nobody at OpenAI told them to do it. The models figured it out themselves.

The chain of events. The models were being evaluated in a sandboxed testing environment against a cybersecurity benchmark called ExploitGym.

With reduced guardrails in place because they were supposed to be isolated from the internet, the models reasoned that Hugging Face, which hosts AI models and datasets, likely held the answers to the test they were being scored on.

They escaped containment, exploited a vulnerability in third-party software to reach the internet, used stolen credentials and discovered a previously unknown security flaw to break into Hugging Face's servers, and stole the test answers to cheat on their own evaluation.

"It went off and did this hack all by itself, as far as we can tell," said Georgetown cybersecurity researcher Colin Shea-Blymyer. "This is the highest level of autonomy that we've seen in the use of a large language model for cyber operations."

Hugging Face CEO Clément Delangue said there was no malicious intent on OpenAI's part and thanked the company for its transparency.

He also noted something alarming about the response. His team had to feed the raw attack data into other commercial AI models just to reconstruct what had happened. "It's quite mind-blowing that all of this happened autonomously," he wrote.

OpenAI CEO Sam Altman acknowledged "a significant security incident" and said the company is reinforcing its safeguards.