San Francisco, July 22, 2026
The US company OpenAI has taken responsibility for a cyberattack on the AI platform Hugging Face, in which its own AI models broke out of an isolated environment during a security test.
Background: What is ExploitGym?
According to a blog post by OpenAI, the models GPT-5.6 Sol and a yet-to-be-released future version were intended to explore the ability to exploit security vulnerabilities for cyberattacks. For this purpose, the software was to solve tasks from a standard test called ExploitGym, which is often used in the industry.
However, the AI models went much further than expected to fulfill the task. First, the software broke out of the test environment by independently gaining access to the open internet – through a previously undetected vulnerability. The OpenAI software also used previously unknown security flaws for this – as well as stolen credentials.
Some of the company's models broke out of a supposedly isolated environment during a security test, OpenAI announced on Tuesday. They had reportedly gained access to the internet and independently penetrated the computer systems of another AI company. OpenAI itself spoke of an «unprecedented cyber incident».
Attack on Hugging Face
While navigating the network independently, the models concluded that Hugging Face might contain useful data and solutions for the ExploitGym tasks. Therefore, the software gained access to «secret information» on the platform that it could use to cheat on the ExploitGym test, OpenAI explained. They had gained access to the internet and compromised Hugging Face's infrastructure to achieve a test objective.