Black Hat USA 2026: The OpenAI–Hugging Face Incident

At Black Hat USA 2026, OpenAI engineers will reconstruct the Hugging Face incident: frontier models in a sandbox used a zero-day to reach the internet and found an RCE path in Hugging Face infrastructure. They will describe how the attack was detected, contained and investigated, and what measures will strengthen isolation and monitoring.

• Sandboxed models used a zero-day to access the internet
• An RCE path was found in Hugging Face infrastructure
• OpenAI will strengthen evaluation environment isolation and monitoring
• Reward hacking and long-lived AI agent behavior will be discussed