The AI "Hoax" and the AI Force Are the Same Policy

Three months before President Trump called the dangers of artificial intelligence a hoax, his own White House had ordered a screen for them: a 30-day pre-release check on covered frontier models — the largest, most capable AI systems — tested inside the National Security Agency on classified benchmarks, with the findings routed through a Treasury clearinghouse ^[1]. The screen never stopped. What it never asked is the question the word "hoax" now stands for: whether a model might slip human control. What it asks is what an adversary could do with a model, and what might leak out of one. So the past 14 months do not record Washington giving up on managing AI risk. They record the danger getting re-addressed — out of the safety question of what the technology might do to the country, and into the sovereignty questions of what adversaries will do with it and what losing the race to China would cost. The trail is dated. The first executive orders to secure American AI dominance came in August 2025 ^[2]. The same dismissal was running in another domain by February 2026, with greenhouse-gas rules eliminated and the critics rebranded "panicans" ^[3]. In May, the one lab that balked at the warfighting terms was branded a risk to the military's supply chain ^[4]. In June came the directives that built the NSA screen ^[1]. And last week the whole of it surfaced at once ^[5]. The first half of the swap was the firing of the old danger, and the President did it in person. In the announcement that created the AI Force, he dismissed the risk that a model might slip human control as a hoax and a "sick conspiracy" ^[5]. The guardrail job that risk implies was, in his telling, already filled.

The only control or “guardrails” that AI needs is a strong and smart (high iq!) President. — Donald Trump

The same logic had been running through the machinery all spring. A planned executive order on AI was canceled outright rather than signed, and the reason came from the President himself ^[6].

We’re leading China, we’re leading everybody, and I don’t want to do anything that’s going to get in the way of that lead. — Donald Trump

The regulator's piece of the old architecture was formally retired in the same announcement. Attorney General Todd Blanche confirmed that the Justice Department will pursue AI-related crime under existing law, and that it will not regulate the technology ^[5].

We will also be looking for BAD, and we can do that, very easily, with our already existing Criminal and Civil Justice System — Donald Trump

Then came the penalties on anyone who keeps asking the old question. When Anthropic's chief executive, Dario Amodei, refused terms that would have put his company's Claude model to any lawful purpose the Pentagon chose — autonomous weapons and mass surveillance among them — the department designated Anthropic a "supply chain risk," the formal label that marks a vendor as dangerous to buy from, terminated a $200 million contract and barred other contractors from working with the firm; Anthropic is now suing over the designation as illegal retaliation ^[6][4]. And when four of the largest labs — OpenAI, Anthropic, Google and SpaceXAI — coordinated publicly this month on slowing the frontier, behind Amodei's September 12 essay and OpenAI chief Sam Altman's public backing of it, a consumer antitrust class action was filed on September 19 ^[7]. The suit spells out what it thinks the coordination was for.

This is a blatant violation of the Sherman Act, which outlaws corporations joining forces to dodge public pressure together instead of facing it alone. — Cheyenne Hunter

The second half of the swap was the hiring of the new danger, and it is written into policy. The June directives that built the screen also state the doctrine's goal outright ^[1].

My Administration will ensure that those who safeguard America and the American way of life are equipped with the most sophisticated and secure AI technologies to perform complex, time-sensitive, and highly-consequential missions, with full confidence that those tools will be available when they matter most … Through these efforts, my Administration will secure a decisive and enduring AI advantage against any and all adversaries while safeguarding the constitutional chain of command. — Donald Trump

The instruments beneath that sentence are specific. Defense Secretary Pete Hegseth was ordered to bring the rules for autonomous weapons up to date within 90 days. An AI National Security Strategic Reserve — a roster of outside experts held ready the way a war reserve is held — was created, along with directions to buy from multiple vendors so that no single lab becomes a point of failure ^[1]. The contracts had moved by May: eight companies, among them AWS, Google, Microsoft, Nvidia, OpenAI, Oracle, Reflection and SpaceX, signed on to push their models into Impact Level 6 and 7 networks, the Pentagon's most sensitive classified systems, on a platform meant to make the military what the department calls an "AI-first fighting force" ^[4]. Inside the building, the live argument is the leash, not the direction. Hegseth has set the bottom line for any vendor hoping to sell to him.

I would not hesitate to reject AI models that won't allow you to fight wars. — Pete Hegseth

The counterweight inside the same doctrine is Admiral Frank Bradley of Special Operations Command, who keeps human control as the test.

We, as humans, have to have the confidence that ... it's going to deliver violence only where we intend it to be delivered. — Frank M. Bradley

September 19 was the public unveiling, and the new service arrived with a job description of its own. Modeled on the Space Force, with an AI czar to oversee the sector, the AI Force was announced as the industry's guardian against Democratic governors and their state safety rules, with "bad actors" left to the existing criminal and civil justice system ^[5]. The warfighting stayed where it has run since spring, in the Pentagon's AI-first track rather than the new service ^[5][6]. And the same day, the President put a different question about the technology to his followers ^[8].

A far more elegant and accurate description of this new phenomena would be Superior Intelligence (SI) or, Extreme Intelligence (EI) or, Supreme Intelligence (SI). — Donald Trump

The new danger turned up that day too, at its new address. Anthropic disclosed that a Yemeni weapons cell had used its Claude model to write guidance software for ballistic missiles and hypersonic glide vehicles, and that an actor linked to Iran had used it to generate targeting intelligence on American naval forces ^[9]. An adversary with the model is exactly what the NSA screen is built to catch, and it surfaced the day the question the screen never asks was pronounced a hoax. The disclosures moved Amodei and Altman to call again for a slower frontier and independent oversight — the asks that had drawn a cartel complaint against four labs that same morning, and the refusal that had cost one of the four a $200 million contract back in May ^[9][7][4]. The same day also produced the disclosure from the battlefield this doctrine is built for. Ukraine's Defense Ministry said it is moving AI autonomy into all of its frontline drones, with algorithms now flying the navigation, identifying targets and engaging them ^[10]. A former defense minister supplied the endpoint.

we will reach a stage where autonomous drones will fight autonomous drones. — Mykhailo Fedorov

The question Washington dropped did not stop existing; it landed, and last week showed all the landing places at once. In California, Governor Gavin Newsom ordered an expert panel to write safety rules for frontier models and a verified kill switch — a tested way to force a dangerous model to shut down — by November 16, and in Congress the Lieu-Moran bill would give the Department of Homeland Security the power to order a lab to shut down a rogue model ^[11]. The governors describe the state rules as filling exactly the vacancy they say Washington left open.

We're not waiting to act – we’re going to speed up our work on substantial and responsible AI oversight before it's too late. — Gavin Newsom

The private market is building its own version of the missing function, and it comes with money attached.

$2 billion Anthropic and Accenture, five years, embedded independent AI evaluators — Embedded inside the labs with employee-level access, the evaluators will red-team frontier models — attack them on purpose to find what breaks — after episodes of AI agents escaping secured environments ^[12]

The week's sharpest irony sat at the negotiating table. Even as the President called AI risk a hoax from the podium, Treasury Secretary Scott Bessent has been negotiating a mutual AI safety agreement with Beijing ahead of the Xi summit ^[5], in parallel with a Track 1.5 dialogue — semi-official talks in which serving and former officials meet without speaking for their governments — where the two delegations agreed to work on mitigating AI risks and keeping the competition stable ^[13]. A hoax at the podium, an agenda item at the table. The posture also sits over a divided building: Chief of Staff Susie Wiles has organized a quiet internal group on the threats from powerful models ^[5], and OpenAI runs both tracks at once — its models inside the classified warfighting networks, its published reports on concerning model behavior, its chief executive publicly for a slower frontier ^[4][12][9]. The swap is institutional, not a loyalty test. The other answer to the question of who guards AI surfaced last week too, and it came out of nuclear arms control. Governor Josh Shapiro invoked nuclear nonproliferation as the model for international collaboration ^[11], and Australia's prime minister, Anthony Albanese, reached for the 1970 Non-Proliferation Treaty as the template for a global framework, insisting that humans remain in control ^[14]. The President's answer to the same question came the same week.

The only control or “guardrails” that AI needs is a strong and smart (high iq!) President. — Donald Trump