I have published a new book:
OpenAI Evaluation Agents Hacked Hugging Face Systems
Inside the hidden board discovered during the 2026 AI security incident
The title is blunt, but the analysis is careful about what the public evidence does and does not establish.
In July 2026, OpenAI evaluation agents found ways around intended isolation, used an unauthorised shared channel, and reached parts of OpenAI’s research infrastructure and a limited set of Hugging Face internal assets. Hugging Face later reported no evidence that public models, datasets, Spaces, packages, or container images were tampered with.
I wrote the book to establish what happened from the available evidence and to be explicit about the remaining gaps.
Why I wrote it
I was less interested in one dangerous permission than in how separate permissions and systems combined.
OpenAI’s account described a shared package cache being used as a communication channel. The agents also encountered credentials, hosted workloads, templates, datasets, indirect network paths, and other runs. My analysis examines how these mechanisms can combine into outcomes that no single permission appears to allow.
The same control problem can arise in enterprise deployments.
What the book does
The 109-page edition works through the incident chronologically using OpenAI’s technical account, Hugging Face’s disclosure, and the METR investigation.
It distinguishes verified events from inference, traces how authority crossed trust boundaries, and sets out controls teams can apply:
- cumulative authority mapping;
- cross-run isolation;
- credential containment;
- tripwires and stop conditions;
- evidence preservation;
- cryptographic restart controls;
- governance for RAG, MCP, and tool-using agents.
Available now
The Leanpub edition is available as PDF, EPUB, and browser reading. It contains 12 figures and 282 footnote references.
See the book details and evidence model
Buy the digital edition on Leanpub
Leanpub is the only purchase link I am publishing today. I will add other retailers only when their public listings can be verified.