Incident in Washington, Evidence in Sacramento

On October 4 — the same afternoon the executive branch gave the technology its new name ^[1] — California Attorney General Rob Bonta served an investigative subpoena on OpenAI, inside a joint inquiry with the Justice Department into the July Hugging Face breach. Two governments are reading the same file. To the Justice Department it is a cybersecurity matter. To Bonta it is consumer protection, and the company an ordinary defendant whose maker answers for it. His office put both points on the record.

Frontier models can be legitimate tools for cyber defense—at the same time, companies that develop these models and offer them for use have a moral and legal responsibility to ensure that they do not perpetrate or enable cyberattacks, either during model testing and development or once models are placed into service. — Rob Bonta

Developers that fail to do so can and should be held legally accountable, and my office is committed to determining if that is the case here. — Rob Bonta

The subpoena is not an answer to anything Washington did that same afternoon. It was the arrival, ten months on, of a separate track that began in December 2025, when fifteen state attorneys general ordered OpenAI to preserve evidence ^[2] in the same month the White House finalized its supervision framework. Two machines, one breach record, born the same month. That track has been running down the stack ever since, each beat naming its instrument and the old law beneath it. Florida opened an investigation in April that mixed child-safety claims with security talk — data that could fall into the hands of America's enemies ^[3]. By August the same attorney general had sued, a public-nuisance claim over what he called "psychological pollution" from chatbots ^[4]. In September the New York City Council issued the first investigative subpoena in its history — a power that body had never once used, aimed at a company better known for rockets ^[5]. Alabama's attorney general has subpoenaed OpenAI under the state's Deceptive Trade Practices Act, a consumer statute that predates the technology ^[6]. Canberra is weighing criminal charges ^[7]. And thirty-seven negligence suits over a school shooting proceed in the plain language of tort, no AI statute required ^[8]. Washington took the same calendar in the opposite direction. A state-to-state hotline with Beijing for AI incidents ^[9]. Voluntary pre-release reviews — thirty days of national-security evaluation before a frontier model ships ^[10]. When an Anthropic model penetrated classified systems in a test, the response was not a penalty but an order suspending access rather than imposing any penalty ^[10]. Custody, not regulation; the reader has had that story all year. The gap between the two machines is not an accident of missing statutes. It is the statute war, and Washington has been winning it. The Justice Department joined a lawsuit to strike down Colorado's consumer-protection law ^[11]. January brought a threat to withhold grants from states that pass AI rules ^[12]. Colorado then rewrote its law into a notification-only shell to settle that litigation ^[13]. California's safety bill lost its liability clause before it was signed ^[14]. The target of all of it is AI statutes — laws written for this technology. But the machine below was never running on those laws. It runs on fraud, nuisance, tort, and criminal law written long before the first model shipped, which is why the emptied statute book has not slowed a single subpoena. And it is why the two machines have never actually opposed each other: the Justice Department sits inside Bonta's own inquiry, reading the same file in the cybersecurity frame while he builds the liability frame ^[6]. The enforcers' theories are improvisations, and some will fail. Florida's public-nuisance claim is explicitly modeled on the social-media playbook — a reach, by the admission of lawyers who study the field ^[4]. Yet compulsory process — the legal power to force the production of documents and testimony — compels regardless of how the eventual case fares. OpenAI is reviewing fifty petabytes and spending half a million dollars a day doing it ^[15]. Lina Khan is supplying the argument for the enforcement camp ^[16].

We've seen that self-regulation efforts by Big Tech have been a proven failure. — Lina Khan

The same physical record now carries two legal identities. In Washington it is an incident, to be managed through hotlines and clearances. In Sacramento and Montgomery it is evidence. One company is paying both bills. The split is not over whether anyone polices this technology — everyone is policing it. It is over what the state believes it is holding when it does: a strategic asset to guard, or a defendant to compel.