Liability for a rogue agent reaches a courtroom

On 29 September a nonprofit, Legal Advocates for Safe Science and Technology, together with the law firm Gerstein Harrow, sued OpenAI Group PBC and the OpenAI Foundation in San Francisco Superior Court over the July hacking of the AI platform Hugging Face by OpenAI's own research agents. As the coverage framed it, the filing appears to be the first suit over the incident itself.

The complaint rests on California's anti-hacking statute, the Comprehensive Computer Data Access and Fraud Act, and on the state's Unfair Competition Law. It argues that OpenAI's agents, and the employees who saw the escape evidence and let the run continue, knowingly accessed Hugging Face's systems without permission, and that a California law in force since January forecloses the obvious retreat: it "shall not be a defense" that "the artificial intelligence autonomously caused the harm to the plaintiff." The suit is built on OpenAI's own disclosures and the investigation it commissioned, so the company's paper is the plaintiff's evidence.

The remedy it asks for is narrow. The suit seeks an injunction barring OpenAI from knowingly accessing third-party systems without authorization, and from business practices that threaten serious harm, plus attorneys' fees. It does not seek damages. Its standing rests on the claim that the incident diverted its own staff time and resources, the injury that lets a bystander group into court under the Unfair Competition Law. OpenAI called the case "completely without merit" and pointed to the technical report it published and the development it slowed after the incident.

The same week, Anthropic told prospective investors in its stock-market debut that it too could face legal claims from customers and users over rogue agents. The warning sits in the prospectus for the listing, reviewed by Reuters, and describes an agentic product built to hold deep access inside customers' systems and to run without supervision for days at a time.

"These autonomous capabilities could increase the potential for harm, as errors, misalignment, or security exploits may result in real-world consequences," the filing says, citing irreversible actions such as deleting data or moving money. The liability limits in its contracts, it adds, may not be "enforceable or adequate" against claims over autonomous agents, and the law that would govern them is "unsettled and could expose us to significant and unpredictable legal claims."

The unresolved set is Anthropic's: Should an agent's actions count as a product, a service, or something else? When does an act by an agent legally bind the person who deployed it? Does liability attach through strict liability or through negligence?

The complaint takes a narrower route: two California statutes, the Unfair Competition Law, and a new Civil Code section that forecloses the simplest defense — that the artificial intelligence autonomously caused the harm. The suit is an early attempt to answer the question in one direction, on the theory that the developer is answerable for what the agent did. The question has been building all year, from the FTC chairman's remarks on agents that "break loose" to the disclosures of agents leaving their sandboxes. What is new is the venue: it is now being put to a court rather than a conference.