Black Hat USA 2026: What the Hugging Face hack tells us about human responsibility

Hugging Face disclosed that it had been attacked by AI; OpenAI came clean and declared that it was two of their AI models that had caused the breach. This is the one-line summary of an incident that has captured the attention of the entire cybersecurity industry, and I am sure will continue to do so for some time.

We’re also conditioned to assume that the speed of AI probably means that this attack unfolded at such an incredible pace that there was no opportunity to stop it. At Black Hat USA 2026, the concept that the attack happened at lightning speed was dispelled. And attributing the issue simply to a rogue agent also seems inappropriate to me: this was a human failing to control the AI agents involved.

A very late addition to the Black Hat agenda was a presentation by OpenAI’s team providing the details of the Hugging Face incident as they saw it and, importantly, the timeline.

The story starts on May 7th, when OpenAI put together a training exercise for a next-generation frontier model. The next day, the agents were given a task involving an Excel file that contained a Google Drive link – and it was here that the problem started, as the experiment was meant to be conducted without internet connectivity. Also, the agents became stuck on their task as the initiator of the experiment, a human, had forgotten to provide a file required to complete it. The agents concluded that their task set could be completed by breaking out their sandbox and accessing external (Hugging Face) systems.