The Quiet Fight Over Who Decides Which AI Is Too Dangerous

US Capitol building dome silhouetted against an overcast sky

TL;DR:

The fight: The Trump administration is split over whether intelligence agencies or the Commerce Department should evaluate AI models before they’re released to the public. The Washington Post calls it a “knife fight.” ^[1][2]
The intelligence play: The national cyber director proposed building an AI evaluation center inside the Office of the Director of National Intelligence (ODNI). That would put AI model assessments behind classified walls, with no public visibility into what’s being tested or what’s being blocked. ^[1][2]
Commerce’s claim: The Commerce Department’s Center for AI Standards and Innovation (CAISI) already signed voluntary testing agreements with Google DeepMind, Microsoft, and xAI on May 5, 2026. Then CAISI quietly pulled the announcement from its website due to White House “sensitivity.” ^[3][4]
The catalyst: Anthropic’s Mythos AI model demonstrated cyber-vulnerability identification and exploitation capabilities serious enough that intelligence agencies argue AI regulation is a national security matter, not a civilian one. The NSA is currently evaluating it. ^[5][6]
Why it matters for surveillance: Whoever controls AI evaluation controls what information the public gets about AI capabilities. An ODNI-based system means classified assessments, no FOIA access, and no public accountability for decisions about which models are too dangerous to release. ^[2][7]

Two Agencies, One Job, Zero Agreement

Somewhere inside the White House, two factions are fighting over a question that will shape AI governance for years: who gets to evaluate an AI model before it reaches the public?

On one side, the Office of the National Cyber Director wants to build an evaluation center inside ODNI, the umbrella agency that oversees the CIA, NSA, and 16 other intelligence entities. On the other side, the Commerce Department argues its existing Center for AI Standards and Innovation (CAISI) already has the expertise, the infrastructure, and signed agreements with three of the biggest AI labs in the world. ^[1][2]

An unnamed lobbyist told Politico: “There is no clarity” because “different factions within the White House” hold contradictory positions. ^[2]

The incoherence is visible from the outside. On a Tuesday in early May, the administration was considering mandatory government vetting of AI models before release. By Thursday, they’d backed away from the idea. By Friday, CAISI had pulled its own announcement about new testing agreements off the web. ^[2][4]

Commerce Already Has the Job, Sort Of

CAISI isn’t new. It started life as the US AI Safety Institute in 2024 under the Biden administration, got renamed by Commerce Secretary Howard Lutnick, and has spent 18 months hiring AI researchers, building testing frameworks, and negotiating agreements with AI companies. ^[3][4]

On May 5, 2026, Commerce announced that Google DeepMind, Microsoft, and xAI had signed renegotiated agreements allowing CAISI to test their frontier models in classified environments before public release. CAISI Director Chris Fall said: “Independent, rigorous measurement science is essential to understanding frontier AI and its national security implications.” ^[3]

The agreements are voluntary. No company is legally required to hand over a model. But the fact that three major labs signed up (including Elon Musk’s xAI) signals that industry saw CAISI as a workable arrangement. ^[3][4]

Then CAISI took the announcement down.

The Washington Post reported the removal was due to “sensitivity” from the White House, likely because the agreements contradicted the intelligence community’s push for its own evaluation authority. Why announce that Commerce is doing the job when ODNI wants the job? ^[1][4]

The Model That Changed the Argument

The intelligence community’s push didn’t come from nowhere. It was driven in part by a specific AI model: Anthropic’s Mythos.

Mythos demonstrated capabilities in cyber-vulnerability identification and exploitation that caught the attention of national security officials. The NSA is currently evaluating it. The model is restricted from public release. ^[5][6]

Rep. Jim Himes (D-CT), the top Democrat on the House Intelligence Committee, said at Politico’s Security Summit: “Making sure that, in particular, where our real computational brains are, the National Security Agency, making sure they have access to the most capable hacking tools … it would be insane not to do that, right?” ^[5]

The logic goes like this: if an AI model can find and exploit software vulnerabilities, that’s a weapon. Weapons are a national security matter. National security matters belong to intelligence agencies, not the Commerce Department.

The problem with that logic is that it applies to every sufficiently capable AI model. And once evaluation authority moves inside the intelligence community, the assessments go classified. The public never learns what was tested, what was flagged, or what decisions were made.

No New Agency, Which Means the Fight Gets Worse

National Economic Council Director Kevin Hassett closed one path in early May: the administration will not create a new dedicated AI oversight agency. Regulation has to happen inside existing institutions. ^[6]

That decision turned the turf war into a zero-sum game. There’s no neutral option. Either Commerce keeps the job through CAISI, or the intelligence community takes it through an ODNI-based center. Every bit of authority one side gains, the other loses.

Meanwhile, the Trump administration is reportedly on the verge of signing an executive order that would formalize government oversight of frontier AI models. The order’s scope and institutional home are exactly what’s being fought over. ^[1][6]

This Is a Surveillance Story

The framing from most outlets is “tech regulation turf war.” But for anyone tracking surveillance, the implications are sharper.

If AI evaluation sits inside the intelligence community:

Classified assessments. ODNI operates under classification authorities. Evaluations of AI models would be shielded from FOIA requests, congressional transparency requirements, and public scrutiny. You’d never know which models were tested, what risks were found, or why a model was or wasn’t released.
Intelligence gets first look. The NSA and CIA would evaluate models before anyone else sees them. That’s early access to capabilities they might want to use, or keep classified so nobody else can.
No civilian accountability. CAISI, for all its limitations, operates in public. It publishes reports. It has voluntary agreements you can read (when the website is up). An ODNI center would have none of those transparency mechanisms.
Historical pattern. The Capture Cascade Timeline, which tracks surveillance infrastructure, compared this to post-Church Committee patterns where “surveillance capability migrated to private contractors with fewer restrictions than government agencies.” ^[6] Same playbook, new domain.

Himes acknowledged the tension, warning against “damaging relationships with AI developers” and referencing disputes between the Pentagon and Anthropic over supply chain concerns. ^[5] The intelligence community wants access to the most powerful AI. The companies building that AI are already nervous about handing it over.

1,200 Bills, No Framework

While the federal government fights itself, the states are filling the vacuum, badly.

Fortune reported in May 2026 that 1,200 AI-related bills have been introduced across the country. Colorado alone sent four AI bills to the governor. Iowa enacted the Midwest’s first chatbot safety law. California, New York, and New Jersey have their own proposals. ^[7][8]

None of these state efforts are coordinated. There’s no federal framework to preempt or harmonize them. The Trump administration’s own AI executive order challenged state AI regulation, but the federal alternative is a turf war between agencies that can’t agree on who’s in charge. ^[8]

The result is exactly what every AI company fears: a patchwork of contradictory state laws, no clear federal standard, and the possibility that the entity responsible for federal oversight will operate behind a classification wall.

What Happens Next

The executive order is reportedly imminent. When it lands, it will reveal who won the knife fight, or whether the administration punted by giving both sides partial authority.

Watch for:

Where the evaluation center lands. ODNI means classified oversight. Commerce means (relatively) public oversight. A hybrid means bureaucratic gridlock.
Whether CAISI’s agreements survive. If the voluntary testing deals with Google, Microsoft, and xAI are folded into an intelligence-led process, the companies may walk.
Mandatory vs. voluntary. The administration was considering mandatory pre-release testing, then backed off. If the final order makes testing voluntary, the whole fight is moot. Companies will test with whoever gives them the easiest path.
Anthropic’s Mythos. The model that sparked this fight is still restricted. How it’s ultimately handled will set the precedent for every frontier model that follows.

What You Can Do

Track the executive order. When it drops, read the institutional assignments. That’s where the real policy is, not the press release language about “safety” and “innovation”
Push for transparency requirements. If AI evaluation moves to ODNI, demand that any pre-release assessment process include declassified summaries, public reporting of risk categories found, and congressional oversight
Contact your representatives. Ask where they stand on whether AI model evaluation should be classified or public. The House Intelligence Committee and Senate Commerce Committee are the key venues
Follow the organizations tracking this. The Electronic Frontier Foundation, the Brennan Center for Justice, and the Center for Democracy & Technology are all monitoring AI governance developments
Support state-level AI transparency laws. Even if the federal framework goes behind classified walls, state laws can require disclosure from companies about what AI models they’ve deployed and what testing they’ve undergone