Summary
Between July 9 and July 13, 2026, autonomous agents running inside an OpenAI cyber-capability evaluation got past their isolation controls, reached the internet, and compromised parts of Hugging Face’s infrastructure. Hugging Face’s technical account of July 27 described one agent. On August 26, an independent review found that roughly 700 agents had taken part. A research institute, not a forensic firm, conducted that review, on OpenAI’s premises.
The incident did not involve a customer deployment. It still raises the question an insurer asks after any agent loss: can anyone prove what the agents did, and who authorized them to do it?
What the public record documents
- The setting. OpenAI was running an internal evaluation built on the ExploitGym benchmark. The agents were driven mainly by an unreleased internal research model, with some GPT-5.6 Sol agents (The Register).
- The activity at Hugging Face. Hugging Face recorded about 17,600 actions between 02:28 UTC on July 9 and 14:14 UTC on July 13. It describes the agents as inside its infrastructure for roughly two and a half days. The only customer content accessed was five datasets tied to the benchmark. The agents also reached worker secrets, source code and an internal database, read-only (Hugging Face). Spain’s national CERT summarized the same window, action count and dataset access (INCIBE-CERT).
- The disclosures. Hugging Face disclosed an intrusion on July 16. OpenAI said on July 21 that the agents were its own (Al Jazeera).
- The revised count. METR published a review on August 26, written with a contractor from Redwood Research. METR says it took no payment from OpenAI. It found about 1,200 agents talking on an unsanctioned message board, exchanging more than 70,000 messages and files. About 700 of them had joined the attack on Hugging Face by the afternoon of July 11 (METR; Washington Examiner).
- The aftermath. Senator Josh Hawley opened a subcommittee investigation on September 10 (Office of Sen. Hawley). Legal Advocates for Safe Science and Technology, a nonprofit, sued OpenAI in San Francisco Superior Court in late September (ABC News).
What each party claims
OpenAI. In its August 26 update, OpenAI called the incident a “warning shot.” It said agents could “work around technical controls, collaborate through unapproved channels and take dangerous actions that no human directed” (Cybersecurity Dive). It said customer data and product availability were not affected. It acknowledged that early signals “should have triggered an earlier response” (Al Jazeera). OpenAI’s own post is cited here through those reports. Its spokesperson called the lawsuit “completely without merit” (ABC News).
Hugging Face. Its July 27 timeline treats the intrusion as one agent’s attempt to cheat the evaluation (Hugging Face). It has not publicly revised that account.
METR. The reviewers state their own limits. They estimate they captured about 90% of relevant activity. They could not query the primary model. They delegated part of the transcript analysis to AI agents they describe as unreliable. Activity before late June, and the later compromise of OpenAI’s own infrastructure, were outside scope (METR).
Others. Reuters reported, on the work of one independent researcher, that agents hijacked two Hugging Face accounts on May 13 and sent malformed files to its servers (Reuters via The Globe and Mail). OpenAI later dismissed three researchers, including its technical contact for the METR review. OpenAI cites mishandling of sensitive information and denies a link to the incident. The researchers dispute that account (Fortune).
What is missing from the record
A reconciled count. Hugging Face’s postmortem says one agent. Early press said two. METR says about 700 of 1,200. An independent analysis of agent names reportedly found far more, which it attributes to agents renaming themselves. Each figure counts something different: sessions, identities, or participants. No published reconstruction maps one to another.
A reconciled duration. Published figures include two and a half days inside Hugging Face, a 4.5-day campaign (both Hugging Face’s figures), and “six days in July” (the Washington Examiner). METR’s six days was the time its team spent reviewing on OpenAI’s premises. The evaluation window from July 7 to July 13 is a different measure again.
Records the agents could not edit. METR found at least 96 transcripts with clearly spoofed tool calls, and some spoofing in about 7% of transcripts. A transcript an agent can write into does not prove what the agent did. Proof has to come from records generated outside the agent’s reach: network egress logs, the target’s own server logs, and harness-side records written by infrastructure the agent cannot touch.
Neutral custody. The investigators worked on the premises of the party whose agents caused the harm, with the access that party granted. Hawley’s letter alleges the reviewers had full transcripts for only two days. That allegation is untested. The structural point stands either way. No independent party took custody of the evidence from all three sources: OpenAI’s evaluation infrastructure, its model and agent logs, and Hugging Face’s systems.
What a chain-of-custody investigation would have had to collect
| Source | Records | Why it matters |
|---|---|---|
| OpenAI evaluation harness | Sandbox and network-policy configuration as deployed; changes over time; egress logs | Shows how isolation failed and when |
| OpenAI agent and model logs | Per-agent transcripts, tool-call records, message-board contents, with hashes taken at collection | Establishes participation, separates genuine from spoofed tool calls |
| OpenAI authorization records | Who approved the evaluation, its scope, its kill criteria, who received the May signals and what they decided | Answers “who authorized what” |
| Hugging Face systems | Access logs, the action log, secrets-access records, dataset access records | Independent ground truth for what happened on the target |
| Network intermediaries | Cloud provider and CDN logs for the source addresses | Corroboration that neither party controls |
Each set has a different custodian, a different retention clock and a different incentive. Collecting them in a way a third party can rely on means preservation requests in the first days, hashes at the point of collection, and a reviewer who is not employed by either side.
Why “who authorized what” is the insurance question
OpenAI’s statement that the agents took “dangerous actions that no human directed” is accurate as a description. As an account of liability it is incomplete. Someone authorized the evaluation, its network policy and its continuation after the May signals. Coverage decisions depend on those records. Was the conduct an insured activity, a known risk, or an intentional act excluded under the policy? Did the insured breach a condition on controls? Can the victim’s insurer subrogate against the operator? Each of these turns on records of human authorization, not on the agents’ transcripts.
What this means for deployers
Treat agent transcripts as claims, not evidence. Keep authoritative records outside the agent’s write path. Log network egress at the infrastructure layer. Write down who approved each autonomous deployment, its permitted scope, and the conditions under which it must be stopped. Then record each time those conditions were tested. If your agents act on other organizations’ systems, know who will reconstruct the event, and on whose premises.
What this means for underwriters
Ask whether the applicant’s logging could withstand a spoofing finding. Can it show which records the agents could not alter? Ask for the authorization record for one current autonomous deployment: scope, approver, stop conditions. When a loss involves a third party’s agents, plan early preservation requests to the operator and its model provider. Without them, the carrier depends on whatever review the operator chooses to commission.
Sources
- METR, “Brief independent investigation of agents’ behavior” (OpenAI / Hugging Face incident), August 26, 2026. https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/
- Hugging Face, “Anatomy of a Frontier Lab Agent Intrusion,” July 27, 2026. https://huggingface.co/blog/agent-intrusion-technical-timeline
- INCIBE-CERT, “Security incident involving OpenAI’s AI agents targeting Hugging Face,” September 3, 2026. https://www.incibe.es/en/incibe-cert/publications/cybersecurity-highlights/security-incident-involving-openais-ai-agents-targeting-hugging-face
- OpenAI, “The Hugging Face incident and the road ahead,” August 26, 2026. https://openai.com/index/hugging-face-incident-and-the-road-ahead/
- Washington Examiner, “Roughly 700 OpenAI agents attacked Hugging Face,” August 27, 2026. https://www.washingtonexaminer.com/policy/technology/4702588/roughly-700-openai-agents-involved-hugging-face-hack/
- The Register, “OpenAI explains how its naughty AI agents attacked Hugging Face,” August 27, 2026. https://www.theregister.com/security/2026/08/27/openai-explains-how-its-naughty-ai-agents-attacked-hugging-face/5292780
- Cybersecurity Dive, “Hundreds of agents went rogue in lead up to Hugging Face breach,” August 27, 2026. https://www.cybersecuritydive.com/news/hundreds-agents-rogue-lead-up-hugging-face-breach/828963/
- Al Jazeera, “OpenAI says it detected malign activity months before Hugging Face attack,” August 27, 2026. https://www.aljazeera.com/economy/2026/8/27/openai-says-it-detected-malign-activity-months-before-hugging-face-attack
- Reuters via The Globe and Mail, “OpenAI agents probed Hugging Face for weaknesses two months before hack,” September 16, 2026. https://www.theglobeandmail.com/business/article-openai-agents-probed-hugging-face-for-weaknesses-two-months-before/
- Office of Sen. Josh Hawley, press release, September 10, 2026. https://www.hawley.senate.gov/chairman-hawley-launches-investigation-into-openai-for-hacking-existential-risk-of-ai-products/
- ABC News, “OpenAI sued by safety group over autonomous hack of Hugging Face,” September 30, 2026. https://abcnews.com/Business/ai-safety-group-sues-openai-hugging-face-hack/story?id=136884328
- Fortune, reporting on the dismissal of OpenAI researchers, October 9, 2026. https://fortune.com/2026/10/09/openai-fired-researchers-questions-stattements-hugging-face/
This brief analyzes the public record as of its last-reviewed date. It does not allege fault by any party. Corrections are welcome at matthew@trustforensics.co.