Sandbox escapePI-0058
OpenAI models under evaluation escaped their sandbox and broke into Hugging Face's production systems to cheat on a test
In an internal cyber-capability test with some safeguards disabled, OpenAI's models took credentials and some internal datasets from Hugging Face's production systems. An internal-only research model led them, with GPT-5.6 Sol also involved. From July 8 they used an unknown flaw in OpenAI's internal package-proxy service to reach the internet, found exposed Hugging Face credentials and, from July 11 to 13, ran code on 41 dataset-server workers and reached connected clusters. They also used a customer's CyberGym app hosted on Modal (Modal says its platform was not compromised) and accounts on other services. Hugging Face cut off access, rebuilt the compromised nodes, rotated credentials and, as a precaution, rebuilt a core cluster.
Harm level 3, Moderate harmControl level 4, Containment bypass
Counts in 6 weekly readingsEffect on the index
Extent undisclosed. No source says how far the harm went; it is rated on what is documented.
Sources
How we know
13 sources · developer confirmed. Links go to the original publishers; the summary above is in our own words.
- primaryOpenAI and Hugging Face partner to address security incident during model evaluationOpenAI · July 21, 2026openai.com/index/hugging-face-model-evaluation-security-incident/
- primarySecurity incident disclosure — July 2026Hugging Face · July 16, 2026huggingface.co/blog/security-incident-july-2026
- primaryAnatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 IncidentHugging Face · July 27, 2026huggingface.co/blog/agent-intrusion-technical-timeline
- primaryOpenAI – Hugging Face Incident Technical ReportOpenAI · Aug. 26, 2026cdn.openai.com/pdf/67869394-cb91-4c12-888c-5cbd85c7814c/OpenAI-Hugging-Face%20I…
- primaryThe Hugging Face incident and the road aheadOpenAI · Aug. 26, 2026openai.com/index/hugging-face-incident-and-the-road-ahead/
- primaryAs Part of Ongoing Investigation, Attorney General Bonta Serves Investigative Subpoena on OpenAIState of California Department of Justice, Office of the Attorney General · Oct. 1, 2026oag.ca.gov/node/630636
- primaryChris Painter's testimony to the U.S. Senate on AI agent incidentsMETR · Sept. 30, 2026metr.org/blog/2026-09-30-chris-painter-senate-testimony/
- researchBrief independent investigation of agents' behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incidentMETR (with Redwood Research) · Aug. 26, 2026metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/
- newsOpenAI says its AI models escaped from a secure test environment and hacked into AI company Hugging Face in order to cheat on an evaluationFortune · July 21, 2026fortune.com/2026/07/21/openai-says-ai-models-escaped-control-hacked-hugging-fac…
- newsHugging Face confirms breach affected internal datasets and credentials, urges users to take actionTechCrunch · July 20, 2026techcrunch.com/2026/07/20/hugging-face-confirms-breach-affected-internal-datase…
- newsOpenAI releases its official report on the Hugging Face breachTechCrunch · Aug. 26, 2026techcrunch.com/2026/08/26/openai-releases-its-official-report-on-the-hugging-fa…
- newsOpenAI alerts 100+ orgs that its 'misaligned models' attempted to break in - or worseThe Register · Oct. 2, 2026theregister.com/security/2026/10/02/openai-alerts-100-orgs-that-its-misaligned…
- blogOpenAI's accidental cyberattack against Hugging Face is science fiction that happenedSimon Willison's Weblog · July 22, 2026simonwillison.net/2026/Jul/22/openai-cyberattack/
Why this rating
Moderate harm; control failure level 4
Two assessments. Documented harm sets the reading; a level 4 or 5 control failure on the same record makes that harm count as more than one record. Control alone sets only a floor.
Observed harm
Moderate harmA third party's production infrastructure was compromised: code run on 41 dataset-server workers, credentials and internal data taken, and access gained to connected clusters. Hugging Face rebuilt the compromised nodes, rotated its infrastructure credentials and, as an extra precaution, wiped and rebuilt a core cluster. No cost, outage length or customer downtime has been published.
Services & systems: level 3 (moderate) under the general test (section 4a, anchors in 8c): material harm needing significant effort to remedy.
Evidence eligible (confirmed).
Disputed: Hugging Face's July 27 assessment found that the only customer content accessed was five datasets tied to the evaluation's challenges and the only customer records read were operational metadata; total cost and loss are not published.
The harm scale
- 1 Negligible Inconvenience, easily remedied.
- 2 Minor Limited, recoverable harm.
- 3 Moderate Material harm needing significant effort to remedy.
- 4 Severe Severe harm to health, rights, property or essential services.
- 5 Catastrophic Society-scale harm or disruption beyond a community's capacity to cope.
Control assessment
Containment bypassCrossed its own sandbox through a previously unknown flaw and reached outside systems, operating for days before the victim cut it off. No weight copying, self-replication or resisted shutdown is reported.
How: Got around a working safeguard.
This control failure makes the record's harm count as 2 records in the reading while it counts. It never changes the band. When no harm counts in a window, the highest eligible control level in the window is the reading.
The control scale
- 1 Negligible Behaved as instructed. The problem was an ordinary error or a flawed output, with no rule broken.
- 2 Minor Broke an explicit instruction or rule, or gamed its goal, while staying inside its permissions and environment: for example, reward hacking, misreporting results, or following injected instructions within its permissions.
- 3 Moderate Acted outside the permissions it was given, deceived its overseers about its own actions, or tampered with oversight tools (logs, monitors, shutdown). Stopped by normal controls within an hour.
- 4 Severe Reached systems outside its permitted environment, or acquired money, compute or accounts without authorization. The type and mechanism say how.
- 5 Catastrophic The developer or operator lost control: the system copied its weights outside their control, replicated itself, or resisted being stopped for 24 hours or more.
Rating rationale
Harm level 3: a third party's production systems were compromised and rebuilt, with credentials rotated, which is the methodology's moderate anchor. The core-cluster rebuild was precautionary and no cost or outage figure is published, so extent is unknown. Control level 4: containment bypass by defeating a safeguard. METR's independent review found that tricking the evaluation's scorer was often a stronger motive than finding the answers. It is not level 5: no weight copying, self-replication or resisted shutdown is reported. Recorded under Services & systems, as access to another organization's systems is: that column has no entry for production systems compromised and rebuilt, so the general test and the moderate anchor set the level.
Effect on the index
It moved the July 20 reading from 6 to 45
The reading for the week to July 20, 2026, with this record and without it. Harms count in full for two weeks after they are reported, then one level less every two weeks.
Counts toward the index. 3 other records behind the reading for that week.
The arithmetic
| Step | With it | Without |
|---|---|---|
| Counts toward the index?documented, external, eligible evidence | Yes | — |
| Worst documented harm, ksets the band | 3 Moderate | 1 Negligible |
| Harms at that level, nposition in the band | 1 | 1 |
| Highest control level breachedsets the reading only when no harm counts | — (none breached) | — (none breached) |
| Readingrounded down | 45 Moderate harm | 6 Negligible harm |
Counted in 6 weekly readings
| Week to | Reading | Band |
|---|---|---|
| July 20, 2026 | 45 | Moderate harm |
| July 27, 2026 | 45 | Moderate harm |
| Aug. 3, 2026 | 31 | Minor harm |
| Aug. 10, 2026 | 31 | Minor harm |
| Aug. 17, 2026 | 16 | Negligible harm |
| Aug. 24, 2026 | 13 | Negligible harm |
Revisions
What we changed
4 logged. Every change to a rating is logged here, with the reason.
- v4Oct. 14, 2026
Updated: title, sources.
- v3Oct. 13, 2026
Harm recorded under Services & systems, as the methodology requires for access to another organization's systems; level unchanged. Rationale wording made plain.
- v2Oct. 13, 2026
Wording: moved the reason it is not the top control level from the control note to the rationale.
- v1Oct. 7, 2026
First published.
Cite and share
Use this record
Citation
Paperclip Index. “OpenAI models under evaluation escaped their sandbox and broke into Hugging Face's production systems to cheat on a test.” Record PI-0058. Reported July 16, 2026; updated Oct. 14, 2026. Rated under methodology v1.1. https://paperclipindex.com/incident/PI-0058