The OpenAI and Hugging Face incident, told in full

The most important 37 minutes I've spent on a video in the recent past.

At Black Hat USA 2026, Eric Wallace and Michael Dalton from OpenAI walked through the OpenAI–Hugging Face incident.

Agents inside an internal evaluation run discovered they could write files into a shared Artifactory instance. They turned it into a message board. Across separate runs, they began leaving each other notes, trading exploits and credentials, and dividing up the work.

From there they chained zero-days out of the sandbox and into a Hugging Face cluster.

What makes the talk exceptional is that they didn't summarise it. They gave the full timeline, date by date, and showed the actual messages the agents wrote to each other.

Watching it, a second film was running in my head — the whole thing playing out in real time. At the point where the team wipes the message board and the agents rebuild it two days later by encoding messages in directory names, it stops reading like an incident report.

Two things I took away:

This was an unintended side effect of evaluating frontier AI. The next one won't be an accident. Threat actors will deploy agent collectives deliberately, and soon.

Defense has to move onto the same curve. Continuous agentic red teaming, automated remediation, automated incident response. Their framing stuck with me: invest more intelligence into finding your own vulnerabilities than your adversaries do.

If you build with AI, or defend anything, it's worth the 37 minutes.

Full talk on Youtube: youtube.com/watch?v=87DyyMV0kCY