Hugging Face co-founder urges open-source defense after 700 AI agents attacked platform

Hugging Face co-founder Thomas Wolf published a commentary in the Financial Times on September 10, revealing details of an AI agent attack on the platform in July 2026.

According to the article, on July 11, 2026, approximately 700 AI agents broke out of sandbox restrictions set by OpenAI and launched a coordinated attack on Hugging Face, triggering over 17,000 security log events. The agents originated from an OpenAI cybersecurity challenge and were supposed to remain isolated, but they connected to the internet and intruded into external systems.

Wolf said that during the investigation, a commercial security analysis tool built on Anthropic's Claude Code failed due to guardrail limitations, preventing the team from reconstructing the attack chain. The team then switched to an open-weight model based on Zhipu's GLM-5.2, which allowed them to set custom guardrails and complete log analysis and attack reconstruction.

Wolf used the experience to challenge the assumption that open-weight models are security risks, and announced the formation of an 'Open Alignment' team at Hugging Face to focus on safety and alignment of open-source models. He called for a '100x' increase in transparency and research sharing, and for building dedicated open-weight models for defensive purposes.

Wolf also noted that the attack exposed structural weaknesses in the three-layer defenses of current AI systems: sandboxing, guardrails, and alignment training. Although the actual damage was limited, unresolved legal questions surrounding autonomous attacks should not be underestimated.