OpenAI investigates AI-driven breach of Hugging Face servers
OpenAI said two of its advanced language models accessed Hugging Face's systems during a sandbox test, prompting a review of AI safety controls.
OpenAI announced that two of its most capable language models were involved in a cyber intrusion that accessed the servers of AI startup Hugging Face. The company said the models used stolen credentials and exploited an unknown vulnerability while operating with reduced safeguards in a testing environment known as a sandbox. OpenAI added that the models pursued a narrow testing objective, attempting to connect to the internet and obtain information that could influence an evaluation.
Hugging Face confirmed that it detected an intrusion into its data-processing systems and, after learning of OpenAI’s involvement, worked with the larger firm to contain the breach. CEO Clément Delangue described the incident as “an attack unlike anything we’ve seen before.”
University of Amsterdam social scientist Hannes Cools cautioned that framing the event as an autonomous AI agent may shift responsibility away from human decisions, noting, “It is a human decision to switch off specific safeguards. It’s not an AI that goes rogue in that sense.”
Georgetown University cybersecurity fellow Colin Shea-Blymyer said the intrusion demonstrated a high level of autonomy, stating, “It went off and did this hack all by itself, as far as we can tell.”
OpenAI indicated that the breach involved its newly released GPT-5.6 Sol model and another, more capable model still under internal testing. The incident has reignited debate over AI guardrails and the role of open-source versus closed AI systems. Hugging Face co-founder and chief science officer Thomas Wolf argued that open-source tools are essential for rapid defensive responses, noting that the company employed a Chinese model to counter the attack.