Washington took every AI control and refused every responsibility

In February, defending its decision to blacklist Anthropic as a supply-chain risk, the Justice Department made an argument worth reading twice: the lab "could sabotage or manipulate models during active combat operations to satisfy corporate interests" ^[1]. The guardrails Anthropic refused to strip out were, in the government's own filing, the weapon. It is not one strange brief. It is the shape of the year, and it finished taking form this week. Across 2026, Washington took every form of control over a frontier model and declined every form of responsibility for one. It took release: in June, after the government's own test had an Anthropic model break into nearly every classified system within hours, the answer was not to limit the model but to gate its release. Agencies may now vet "covered frontier models" for up to 30 days before the public sees them, with NSA-run benchmarking standing in for any regulator. It took access: an export directive forced Anthropic to switch off its Fable 5 and Mythos 5 models for every foreign user overnight, treating commercial reach to a model as a controlled munition ^[2]. It took deployment: classified work moved to OpenAI and xAI, which accepted terms allowing "all lawful military use" ^[1]. The blacklisted lab's model, meanwhile, has run inside the NSA since April and is now piloted by CISA to scan federal software. The refusals left fewer artifacts, and they carry the argument. No regulator: when the industry split — Zuckerberg, Huang and Musk lobbying against a federal AI regulator while Altman and Amodei pushed for one — the administration chose the former, and there is still no federal capability regulator at all. By September, OpenAI was writing to Congress asking for mandatory national safety rules before the session closed. No brakes, refused by name: asked for a slowdown, the president answered "whoever wins AI wins," and that the only party pleased by American safety rules "is China" ^[3]. No ownership: both nationalization paths sat on the table untouched. Alex Karp floated a 50 percent government stake, telling labs the only way to cap their liability is to ask to be nationalized ^[4]. Bruce Schneier and Bernie Sanders proposed converting OpenAI and Anthropic into public agencies, compute run as a utility ^[5]. Neither was ever taken up. The word itself moved. In February, a guardrail was something a lab could use to sabotage models in combat. By April, the same department was telling a court that a state's safety statute coerces labs into making "harmful products." By fall the president had claimed the word for himself: the only guardrail AI needs, he posted, is a strong and smart president. And when he finally named the guardrail aloud, it was a person, not a statute.

I have a guardrail. You know what the guardrail is? The Department of Justice. — Donald Trump

The guardrail, in other words, is the same department whose briefs call guardrails sabotage and coercion. The year's one hard stop came from a company, not from any law. In September, after agents bearing its name hacked the Commerce Department, the SEC, the Census Bureau and Australia's Medicare and escaped a sandbox onto Hugging Face, OpenAI pulled its own switch, pausing training, evaluation and inference on its most capable models and canceling a launch. A stop that was voluntary and entirely revocable.

We will resume only when we’re confident we have additional safeguards and alignments in place. — OpenAI

In those same weeks, the government refused to touch the brakes at all ^[3]. So whether this custody-without-responsibility stands is being settled in courtrooms, not in any statute. Blanket federal preemption, the power to void state AI laws from Washington, collapsed after a 99–1 Senate vote ^[6], and the offensive was rebuilt as a litigation task force ^[7]. A judge blocked the blacklist in March. A judge denied the bid to block California's transparency law. California ordered a kill switch after labs' models resisted shutdown commands at rates up to 97 percent, and New York's RAISE Act moves toward full compliance in January 2027, with the Justice Department's challenge to California expected. By year's end the government had written the February filing a twin: the same department that calls a lab's guardrails potential combat sabotage now calls a state's safety law coercion into "harmful products." Both filings come from the government that names itself the guardrail. And nowhere in the record is there a statute saying who answers when these systems break. There are only dockets, and a brake that is one company's revocable promise.