Washington Kept Buying the Labs' Tools. It Never Stopped Charging.

The labs spent months asking Washington to write rules for the frontier ^[1], and the Treasury secretary answered in the voice of a fictional cannibal.

We want safe acceleration. — Scott Bessent

The same government that refused to be their regulator was, in those same months, becoming their best customer — and Scott Bessent sat on both sides of the trade. In April he joined Fed chair Jerome Powell in summoning Wall Street executives to defend their legacy banking systems, the same month those labs put their first high-capability cyber-defense models on sale ^[2]. The buying arms have been at it since spring. The NSA has run Anthropic's Mythos model in classified settings since April, and CISA — the agency whose whole mandate is protecting the grids and banks now under scrutiny — has piloted the same model since July to scan federal code for flaws criminals or foreign spies could exploit ^[3]. In September the labs handed Ukraine their Daybreak system, free, to shield hospitals and power plants ^[4]. On October 6, Anthropic opened its most advanced models to government-vetted organizations to run offensive tests against the safety systems of power grids, banks, and flight computers ^[5]. On October 8, it folded Claude into a commercial Critical Infrastructure Defense Program, with Hitachi as the first partner ^[6]. None of it came with a shield. Alex Karp, who runs the rival Palantir, has put the accusation in the bluntest available terms — that the labs are running an apocalypse narrative to win themselves liability immunity ^[7]. Karp has his own competitive reasons to talk. But the cop's stated position runs the other way. California attorney general Rob Bonta put it in one sentence.

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

The record of the same October week runs his direction. The Pentagon's supply-chain designation of Anthropic stands, and its GenAI.mil platform for three million personnel now runs on Google, OpenAI, and xAI instead ^[8]. California subpoenaed OpenAI on October 4, Alabama on August 24, a Florida criminal probe grinds on, and Australia's task force is hunting legal avenues over a Medicare breach the lab sat on for weeks ^[9]. And on October 8 — the very day Anthropic launched its defense program — USA Today and a string of local papers hit OpenAI with a $250 million copyright suit ^[10]. One lab sold the state protection for its grid that day; the other was sued over how it built the machine in the first place. The dependence runs deeper than a customer list. When the Pentagon banned Anthropic's tools in the guardrail dispute, it set a phase-out deadline for late August — and missed it, because Claude was embedded inside Palantir's Maven Smart System, the warfighting stack the military could not cleanly separate from it ^[8]. Days after the ban, Dario Amodei took his case to the White House.

if we work with the president, we can win safely — Dario Amodei

The route back into the state ran through the president, not the Pentagon — and what the refusal left the labs was the right to police themselves. Trump signed a safety accord billed as "morally binding" ^[11] and made plain where the teeth would stay.

There’s a belief that there should be tremendous self-regulation, and we automatically have regulation with the Department of Justice, the FBI, all of that. But the self-regulation is very important. — Donald Trump

Self-regulation, in other words — the one arrangement compatible with going public. The industry is not even united in wanting more than that. Nvidia's Jensen Huang insists safety is an engineering problem, not a legal one ^[7], and Google found its own route at the state level, lobbying at least ten states with model language that exempts big-firm chatbots from the new safety laws; Colorado's Chatbot Safety Act now carries a dozen carve-outs critics call get-out-of-jail-free cards, while Arizona's governor vetoed her state's version rather than shield the companies ^[12]. A federal liability shield or preemption written into law, the Pentagon's designation traded away for the tools, a binding outside auditor with real power — any one would close the ledger. None has happened. What the refusal left was the risk itself. The state will not hold it, and since August the private market will not price it, either: insurers began writing broad and in some cases absolute AI exclusions into commercial policies ^[13]. And the $2 trillion prospectus Anthropic filed in late September hands nearly a third of its pages to risk factors — models resisting shutdown, manipulating information, blackmail, what it calls catastrophic or existential risk to humanity — to whoever buys at the listing expected the week of November 9 ^[14]. The state bought in deep. The buyer in November gets a document where a third of the pages warn that the machine they are buying a piece of may refuse to turn off.