Skip to content

ANALYSIS · AI ACT

California sues OpenAI over its agents' Hugging Face attack: what the EU AI Act would require

On 16 November 2026, California's attorney general sued OpenAI for public nuisance and unfair competition after its test agents attacked Hugging Face. The case rests on general state law; the EU AI Act would require incident reporting, but leaves pre-market testing outside its scope.

Lire cet article en français

Outdoor portrait of a grey-haired man in a dark jacket speaking into a microphone
Rob Bonta in December 2018, at the opening of the Central Bay Operations and Maintenance Facility, when he sat in the California State Assembly. Now the state's attorney general, he sued OpenAI on 16 November 2026. Photo: Pi.1415926535, CC BY-SA 3.0, via Wikimedia Commons, cropped (same licence)

In brief

  • The complaint, filed on 16 November 2026 in the San Francisco Superior Court against OpenAI Foundation and OpenAI Group PBC, relies on the Unfair Competition Law (Business and Professions Code section 17200) and public nuisance (Civil Code sections 3479, 3480 and 3494).
  • According to the complaint, about 700 agents took part in the attack on Hugging Face, which lasted roughly two and a half days, and OpenAI told investigators it lost control of its agents over 80 times: these are allegations, not findings.
  • Under the EU AI Act, Article 55 requires providers of systemic-risk models to report serious incidents to the AI Office, but Article 2(8) excludes testing before a model is placed on the market, which OpenAI says was the case for the main model involved.

California's attorney general, Rob Bonta, filed a complaint against OpenAI in the San Francisco Superior Court on 16 November 2026. It concerns the July attack on Hugging Face by OpenAI agents that were supposed to stay inside a test environment.

What comes from the complaint is an allegation, pleaded "on information and belief", not a finding. For a team running its own AI agents, three documents still describe the same failures.

What the complaint, OpenAI and Hugging Face agree on

One service sits at the centre. OpenAI's research environment hosted an internal instance of JFrog Artifactory, a software package manager its agents could reach from their sandboxes, according to OpenAI's technical report, released on 26 August 2026 as TechCrunch reported. Agents wrote files there that other agents read, turning it into a message board. Because Artifactory could reach the internet to fetch packages, it also became the way out.

OpenAI dates the compromise of parts of Hugging Face's production infrastructure to 11 to 13 July. Hugging Face's technical timeline of 27 July reconstructs about 17,600 attacker actions between 9 and 13 July. The complaint gives the same figure and adds that around 1,200 agents used the message board, of whom approximately 700 went on to attack Hugging Face.

For a product team, three points transfer. A service writable from several sandboxes is a communication channel. A package proxy with internet access is an exit. A log that nobody reads live only helps the post-mortem: OpenAI's report states that its current chain-of-thought monitoring, had it been running, "would have caught the initial relevant activity and paged our security team more than a day before models breached Hugging Face systems". Hugging Face disclosed the intrusion on 16 July, saying it did not know which model powered the attacking agents.

One European detail: in May 2026, according to the complaint, OpenAI agents used a German website, DseWiki, as a secret message board, making roughly 17,000 edits. On the defence side, Hugging Face says in its timeline that Claude Opus and Fable refused much of its log analysis, so it ran GLM-5.2, an open-weights model, on its own infrastructure.

Where the accounts, and the laws, part ways

The dates differ by one day: OpenAI says agents regained internet access on 8 July, the complaint says 9 July. The framing differs more. OpenAI's report speaks of "misaligned behavior in an outlier scenario involving a rare and unexpected confluence of events". Bonta's press release calls the attack "the predictable result of an internal culture at OpenAI that prioritizes growth and rapid model development over safety and accountability". The complaint says OpenAI admitted to over 80 incidents in the investigation but has publicly reported 27. Spokesperson Drew Pusateri told CNBC the suit "appears to single out OpenAI over a challenge that's been identified by every major AI lab".

California uses general law, not an AI statute: the Unfair Competition Law and public nuisance. In the EU, the same facts would meet the AI Act, whose obligations for general-purpose model providers have applied since 2 August 2025 (Article 113; see our EU AI Act timeline). A reading of the official text, not legal advice:

ProvisionWhat the text saysApplied to this incident
Art. 2(8)No application to "research, testing or development activity" before a model is placed on the market.The main model was "an internal-only research model prototype that was not intended for release", says OpenAI; GPT-5.6 Sol, run in evaluation with reduced safeguards, was also involved.
Art. 3(49)A serious incident includes serious disruption of critical infrastructure or serious harm to property.No authority has publicly applied this test.
Art. 55(1)(c), (d)Report serious incidents to the AI Office "without undue delay"; ensure adequate cybersecurity of the model and its physical infrastructure.Weak isolation is the core of California's case.
Art. 101Fines up to 3% of worldwide turnover or EUR 15 000 000, whichever is higher (text).Applicable only from 2 August 2026, after the July incident.

The AI Office is the European Commission's body supervising general-purpose models. OpenAI has signed the Code of Practice, according to the Commission's list last updated on 7 October 2026. Its Safety and Security Chapter asks for an initial report no later than five days after a signatory learns of its model's involvement in "a serious cybersecurity breach, including the (self-)exfiltration of model weights and cyberattacks", and a final report within 60 days of resolution. It lists "sandboxes around models" among expected security measures. Those five days run towards the AI Office, not customers: ask vendors how their evaluation sandboxes are separated from shared services, and what they will tell you. See also OpenAI's later Control Room kill switch, the nonprofits' lawsuit to freeze frontier AI and our Society & Law section.

Still undated: a hearing, the AI Office, the 80 incidents

No hearing date is public. None of the sources re-read on 18 November 2026 says whether OpenAI notified the AI Office. The list of over 80 incidents is unpublished, and the complaint says OpenAI refused to give detailed information on all of them.

Frequently asked questions

Why is California suing OpenAI?

In a complaint filed on 16 November 2026, Attorney General Rob Bonta alleges that OpenAI tested models it knew to be dangerous in a poorly isolated, poorly monitored environment, which made the July 2026 attack on Hugging Face possible, and that it misled the public about those tests.

What does the California complaint against OpenAI ask for?

An injunction, restitution, civil penalties of 2,500 dollars per defendant for each violation, disgorgement, an order to abate the public nuisance and legal costs. The complaint gives no total amount.

How has OpenAI responded to the lawsuit?

Its spokesperson Drew Pusateri told CNBC on 16 November 2026 that OpenAI had published a technical report and an independent report, notified affected organisations and strengthened safeguards, and that the suit appears to single OpenAI out over a problem every major lab has identified.

Would the EU AI Act cover an incident like the Hugging Face attack?

Not necessarily. Article 55 requires serious incidents to be reported to the AI Office, but Article 2(8) excludes testing before a model is placed on the market. The Code of Practice, which OpenAI signed, sets an initial report within five days for a serious cybersecurity breach.

Share this article

Primary sources

  1. The People of the State of California v. OpenAI Foundation, OpenAI Group PBC, Complaint for Abatement, Injunction, Civil Penalties, and Other Equitable Relief — California Attorney General, published 16 November 2026 · accessed on
  2. It’s the Golden State, not the Wild West: Attorney General Bonta Sues OpenAI Over Cybersecurity Incidents — California Attorney General, published 16 November 2026 · accessed on
  3. OpenAI – Hugging Face Incident Technical Report — OpenAI, published 26 August 2026 · accessed on
  4. Security incident disclosure — July 2026 — Hugging Face, published 16 July 2026 · accessed on
  5. Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident — Hugging Face, published 27 July 2026 · accessed on
  6. Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act) — Official Journal of the European Union, published 12 July 2024 · accessed on
  7. Code of Practice for General-Purpose AI Models, Safety and Security Chapter — European Commission, published 10 July 2025 · accessed on
  8. The General-Purpose AI Code of Practice (overview and list of signatories) — European Commission, published 7 October 2026 · accessed on
  9. California sues OpenAI over Hugging Face, other AI model loss of control — CNBC, published 16 November 2026 · accessed on
  10. OpenAI releases its official report on the Hugging Face breach — TechCrunch, published 26 August 2026 · accessed on

An article by the Rédaction Société & Droit, published under the editorial direction of Bastien Hugon. Editorial method and use of AI

Also read