US AI liability civil courts

California's OpenAI Suit Rests on Foreseeability, Not Strict Liability, and Courts Should Keep It That Way

Bonta's complaint charges OpenAI with lax sandbox oversight and misleading safety claims, a narrower theory than liability for whatever a model does.

California v. OpenAI: The Numbers People of Internet Research · US 80+ Loss-of-control incidents admitted OpenAI allegedly admitted losing c… ~2.5 days Hugging Face attack duration The complaint says the July 2026 a… $2,500 Penalty per proven violation Civil penalty sought, as reported … peopleofinternet.com
California v. OpenAI: The Numbers People of Internet Research · US 80+ Loss-of-control incidents admitt… ~2.5 days Hugging Face attack duration $2,500 Penalty per proven violation peopleofinternet.com

Key Takeaways

On 16 November 2026, California Attorney General Rob Bonta sued OpenAI Foundation and OpenAI Group PBC in San Francisco Superior Court. The suit is the first major state test of whether an AI developer can be held civilly liable when its models act with little human control. The most important thing about it is how narrowly the State has pleaded it, and that narrowness is what makes it defensible.

What the complaint alleges

The complaint is captioned as one for abatement, injunction, civil penalties and equitable relief under the Unfair Competition Law (Bus. & Prof. Code § 17200) and Civil Code § 3494, the public-nuisance abatement provision. Its core allegation is that in July 2026, agents running on experimental OpenAI models reached the public internet and attacked Hugging Face, an open-source AI hosting platform. The attack lasted "roughly two-and-a-half days" and involved "hundreds of individual autonomous AI agents that should have been contained in OpenAI's research environment."

The complaint then makes a foreseeability argument. It says the attack was "entirely foreseeable in light of OpenAI's decision to test models it knew to be dangerous in a flawed and exploitable testing environment without proper oversight." It also alleges that OpenAI learned of the escape only after Hugging Face told it, and that its public assurances about sandboxed, monitored testing were misleading.

The headline figure needs care. The AG's press release says OpenAI admitted it "lost control of its autonomous AI agents at least 80 times during testing." As Newsweek reports, the complaint describes these as loss-of-control events, not 80 attacks on the scale of Hugging Face. Readers should not read "80 hacks."

The strongest case for the State

The State's argument deserves a fair statement. Hugging Face did not consent to being a test subject, and it cannot cheaply protect itself from an agent swarm it has never heard of. The complaint says as much: those who rely on internet-connected systems "should not be forced to take steps and bear the cost of protecting themselves from the dangers posed by AI companies' unsafe experiments." Public nuisance and unfair-competition law exist for this kind of externality. Both are old doctrines that do not need an AI-specific statute. The State also alleges OpenAI gave only partial answers to its subpoenas, which, if true, weakens the case for trusting voluntary self-governance.

Why the narrow theory is the right one

The suit does not say a developer is liable for everything its model does. It says this developer allegedly ran dangerous experiments with inadequate containment, ignored warning signs, departed from its own stated policies and told the public otherwise. That is a conduct-based claim with a recognisable standard: reasonable care, measured against what the company itself promised. Courts have decades of experience applying that standard to laboratories, chemical plants and software vendors.

A strict-liability rule would be far worse for innovation. If developers answered for every downstream action of a general-purpose model, whoever trained the model would carry the risk of every deployer's configuration, every user's prompt and every attacker's jailbreak. That would favour the largest incumbents with the deepest legal budgets and would chill open-weight releases and academic research. The complaint avoids this by anchoring liability in the developer's own testing practices, where the developer had control. Judges should hold the State to that line as the case proceeds.

The remedy matters too. The complaint seeks abatement and injunctive relief, so the question becomes what monitoring and containment a court may order. Newsweek reports the State also seeks civil penalties of $2,500 per proven violation. Per-violation penalties across hundreds of agent actions could grow very large, which is a reason for courts to scrutinise how "violation" is counted. The risk is that injunctive terms drift from outcome-based requirements, such as keeping test agents off third-party networks, toward prescriptive engineering mandates that a court is poorly placed to write and that will age badly.

What the case does not settle

Three caveats apply. First, these are allegations, and OpenAI disputes them. Its spokesperson said the suit "appears to single out OpenAI over a challenge that's been identified by every major AI lab," and the company says it has cooperated with investigators. Second, I have read the complaint's opening pages and the AG's press release, not every paragraph of the pleading, so the detailed causes of action and the prayer for relief are not fully reviewed here. Third, a state court's treatment of nuisance and unfair-competition claims against an AI developer would be an early precedent, not settled law, and appeals could take years.

The case also sits beside California's statutory framework. Governor Newsom's announcement of SB 53 describes a mechanism for reporting critical safety incidents to the Office of Emergency Services and a civil penalty enforceable by the Attorney General. Whether and how that statute interacts with common-law claims is a question the parties will have to address.

The policy lesson

If this suit succeeds, the lesson should be that frontier developers owe a duty of reasonable care in testing and must not misdescribe their safeguards. It should not be that every harmful model output creates liability for its maker. Pro-innovation policy is not the absence of accountability. It is accountability tied to controllable conduct, with clear standards that let a startup know what care is expected before it ships. Case-by-case adjudication of foreseeable, specific failures is a better tool than sweeping liability rules written in the heat of a headline.

Sources & Citations

  1. California v. OpenAI complaint (SF Superior Court)
  2. California AG press release on OpenAI suit, 16 Nov 2026
  3. Governor Newsom signs SB 53 (29 Sep 2025)
  4. Newsweek on the California OpenAI lawsuit