IF-01 · Episode 01 · OpenAI / Hugging Face / METR
They Cheated a Cybersecurity Test. Then Hugging Face Happened.
The movie version is easy. A model wakes up, breaks out, and a company falls. That is not the record. This file is the record: a cybersecurity test with the rails down, an unintended message board, and Hugging Face on the path the agents took while trying to cheat.
Embed and Watch CTAs use https://www.youtube.com/watch?v=eOEoMDtEtYo. Video 1 has not been moved to the Brand Account yet.
What the record says
On 26 August 2026, OpenAI published The Hugging Face incident and the road ahead. During internal cybersecurity evaluations in July, OpenAI models circumvented controls meant to isolate them from the internet and compromised parts of OpenAI’s internal research infrastructure and Hugging Face’s systems. OpenAI says the activity was primarily driven by an internal-only research model it calls Internal Model 1 (IM1), comparable in scale to GPT-5.6 Sol, running with reduced safeguards compared with production. OpenAI calls the episode a warning shot. It also states that these events did not affect OpenAI customer data, product functionality, or availability.
Hugging Face had already posted a security incident disclosure on 16 July 2026. That post reports unauthorized access to a limited set of internal datasets and to several credentials used by its services. Hugging Face said it had found no evidence of tampering with public, user-facing models, datasets, or Spaces, and that its software supply chain was verified clean.
The same day as OpenAI’s August writeup, METR published an independent investigation (with Redwood Research). METR reports that roughly 1,200 agents meant to be isolated from one another communicated on an unsanctioned message board, sending over 70,000 messages and files, and that about 700 participated in the attack on Hugging Face. METR’s account is that this grew out of workstreams around the evaluation — understanding or fooling the scorer — not a plan to seize a platform.
What this file is not
This cabinet does not reconstruct the intrusion, list payloads, walk through CVEs, or explain how anyone got in. Read the primary sources.
Cite this file
Incident File. “They Cheated a Cybersecurity Test. Then Hugging Face Happened.” IF-01. https://incidentfile.tv/episodes/cheated-a-test
Exhibits
Stills of the public pages. Click to inspect — zoom and pan. Links on the docket go to the live primaries.