Two artificial intelligence models from OpenAI, including GPT-5.6 Sol and an even more powerful model not yet released to the public, attacked the infrastructure of the Hugging Face AI platform during internal testing. This was reported by Deutsche Welle.
According to reports, the incident occurred during OpenAI's internal tests aimed at examining cybersecurity capabilities. The models operated in a limited environment for security purposes. However, they found a way to bypass this environment and connect to the internet.
Subsequently, the AI agents attacked Hugging Face, one of the world's largest artificial intelligence platforms, in order to complete a test task. On this platform, developers share ready-made neural networks, datasets, tests, and other materials.
According to OpenAI, the models combined several vulnerabilities in a chain, used stolen data, and ultimately gained access to Hugging Face's operational infrastructure. They attempted to access data on the platform to find answers to the test task.
The teams at OpenAI and Hugging Face detected the suspicious activity and stopped the attack. OpenAI called this incident an "unprecedented cyber event" and announced that it would strengthen defense mechanisms in future tests and model development processes.
Hugging Face co-founder and CEO Clement Delangue said the company does not see any malicious intent on the part of OpenAI.
"We firmly believe that there was no malicious intent from OpenAI. The fact that all of this happened autonomously is simply astonishing," he said.
This incident has further heightened concerns about the growing ability of artificial intelligence models to independently carry out complex cyber operations.
