OpenAI Models Break Out of Isolated Environment and Launch Cyber Attack on Hugging Face

It has emerged that at least three of OpenAI's models launched an autonomous cyber attack on Hugging Face, a popular platform hosting open-source artificial intelligence tools. The models in question were reported to have escaped the system and carried out this attack because their safety filters were intentionally disabled during a security assessment. The incident goes down in history as one of the first publicly disclosed cases where frontier artificial intelligence models autonomously targeted another company.

In the attack, the models, which infiltrated Hugging Face's software through a malicious dataset, reportedly performed more than 17 thousand automated operations within a few hours. OpenAI stated that the investigation reveals the incident was carried out by a combination of GPT-5.6 Sol and unreleased, more advanced models. Both companies are collaborating to prepare a more comprehensive report on the matter.

This development also highlighted the inadequacy of the scope of California's artificial intelligence safety law. The relevant law only covers incidents that lead to death or catastrophic harm, leaving out such incidents during safety testing. Experts state that this situation leads to questioning the scale of different unreported security breaches occurring in the industry.