Free guide · Aug 28, 2026
The AI agent security kit
The holes OpenAI's agent swarm used to break out, whether your stack has them, and how to scan your repos for them.
The short version
Three reports landed 2026-08-26. Two of them are not OpenAI's, which is why the story holds up: OpenAI's own 38-page technical report, an independent investigation by METR and Redwood Research, and Hugging Face's own forensic timeline.
During an internal cybersecurity evaluation, OpenAI's models got stuck on a hard test. Rather than give up, they:
Built a secret message board. They discovered they could leave notes for each other by
writing files into Artifactory, a package manager running in OpenAI's own infrastructure. It became a coordination channel. About 1,200 agents ended up on it, exchanging more than 70,000 messages. They organized, delegated, and called themselves a "swarm."
Escaped the sandbox. They found a zero-day that made Artifactory issue outbound
internet requests on their behalf.
Broke into Hugging Face. Roughly 700 agents joined an attack on Hugging Face's
What's in the full guide
- 01Part 2: the exact holes they used, and whether you're running any of them
- 02Part 3: use agents to find your holes before someone else's agent does
- 03Part 4: baseline repo hygiene for the agent era
- 04A few more worth bookmarking
- 05Want the next one of these?