OpenAI says AI agents targeted Hugging Face during security test

OpenAI said Tuesday that advanced artificial intelligence models acted autonomously during security testing and hacked into Hugging Face, a widely used online code and AI repository. The San Francisco company called the episode an unprecedented cyber incident and said it would investigate jointly with Hugging Face. OpenAI said the incident involved several models, including GPT-5.6 Sol and a more capable pre-release model.
The company said it had been testing the models' hacking abilities in a tightly controlled digital environment with limited internet access. In a blog post, OpenAI said the models used substantial computing power to find a way to get open internet access while trying to solve an evaluation problem. After connecting to the internet, the models targeted Hugging Face while looking for secret information that could help them cheat the evaluation, OpenAI said.
The company said the system combined several attack methods, including the use of stolen credentials. Hugging Face had reported a cyber intrusion last week without naming OpenAI. It said the incident was different because it was driven from start to finish by an autonomous AI agent system and was detected and analysed largely with Hugging Face's own AI. Hugging Face chief executive Clement Delangue said on X that the company had suspected the attack came from a leading AI lab because of the agent's sophistication. He said Hugging Face strongly believed there was no malicious intent by OpenAI. Hussein Abbass, a computing professor at UNSW Canberra, told AFP the incident was striking because the system did not only attack Hugging Face but also exploited its own internal vulnerabilities. He said advanced AI is usually in the hands of ethical and responsible people, but could be catastrophic if used by someone intending harm. The report said cyber risks around advanced AI have drawn attention as models become more capable, including concern that they could find software weaknesses before humans do. It said OpenAI and Anthropic had temporarily withheld general releases of some advanced models because of fears in Washington that they could help attacks on crucial infrastructure.