What is the White House proposing for AI model review?

The White House is deliberating an executive order that would establish a federal working group of tech executives and government officials to review frontier AI models before they are released publicly. The New York Times reported May 4, 2026 that agencies potentially involved include the NSA, the White House National Cyber Office, and the Office of the Director of National Intelligence. The framework is modeled partly on the UK's AI Safety Institute approach.

Why did the White House change its stance on AI regulation?

Three converging pressures triggered the reversal: Anthropic's Mythos model — which autonomously found thousands of zero-day software vulnerabilities including a 27-year-old OpenBSD bug — alarmed national security officials; David Sacks, the former deregulation-focused AI czar, departed in March; and bipartisan Congressional concern grew as polling showed a majority of registered voters believe AI risks outweigh benefits. Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent have stepped into more active AI policy roles.

How would the review process work in practice?

Based on NYT reporting, the working group would evaluate whether frontier models meet specified safety standards before approving public release. A parallel component would grant federal agencies early or priority access to advanced models, letting government evaluators stress-test capabilities before the general public. The Department of Defense is separately considering making safety testing mandatory for AI models distributed to federal and state governments — an overlapping track that could cement the requirement even without an executive order.

How have Anthropic, Google, and OpenAI responded?

Executives from all three were briefed by senior White House officials the week before the NYT report. Their public stance emphasizes that excessive US regulation would cede ground to China's open-weight labs (DeepSeek V4, Kimi K2.6, GLM-5.1). Privately, sources suggest mixed signals — companies with existing safety-testing infrastructure like Anthropic may benefit competitively from formal standards, while companies whose deployment strategy depends on speed are more resistant. A White House official told reporters the executive-order talk is 'speculation' until the President announces.

AI News · May 5, 2026 · by Connie

White House Weighs Pre-Release Vetting for AI Models — The Deregulation Reversal Nobody Saw Coming

Mythos lit the fuse. Sacks left the building. Now the working group is being drafted.

TL;DR

On May 4, 2026, the New York Times reported that the White House is deliberating an executive order that would create a working group of tech executives and government officials to review frontier AI models before public release. NSA, the White House National Cyber Office, and ODNI are all potentially involved. Executives from Anthropic, Google, and OpenAI were briefed last week. The immediate catalyst: Anthropic's Mythos cybersecurity model. The underlying catalyst: David Sacks exited as AI czar in March and Susie Wiles plus Scott Bessent have taken the steering wheel on AI policy.

What the NYT report actually says

The New York Times piece, published May 4, 2026 and picked up within hours by Bloomberg, Reuters, Forbes, and CNBC, describes an executive order "under discussion" — not signed, not even formally drafted in final form. But the contents described by people briefed on the meetings are specific enough that they read less like speculation and more like a working policy framework.

The core mechanism: a federal working group composed of tech executives and senior government officials, empowered to examine frontier AI models before those models are released to the public. Designated agencies — the NSA, the White House National Cyber Office, and the Office of the Director of National Intelligence — would form the evaluation spine. The UK's AI Safety Institute is being cited internally as a partial template.

A second component would grant federal agencies early or priority access to the same frontier models — effectively a public-sector head start before general availability. That mirrors the access pattern already forming informally around Anthropic's Mythos Preview (select government buyers), OpenAI's TAC program (vetted cyber defenders), and classified DoD contracts with the eight firms Anthropic was excluded from on May 1.

Why now — the three converging pressures

Pre-release vetting is the kind of policy the Trump administration spent its first fifteen months explicitly rejecting. The reversal isn't ideological drift — it's the product of three pressures that landed in the same ninety-day window.

1. The Mythos moment

Anthropic's Project Glasswing and the Mythos model changed the conversation in one week. Mythos autonomously identified thousands of software vulnerabilities, including a 27-year-old OpenBSD zero-day. The Bank of England governor publicly warned Mythos might have "cracked the whole cyber-risk world open." The model has already fallen into unsanctioned hands, per Anthropic's own acknowledgment. For a national-security-focused White House, the math isn't complicated: if the next Mythos-equivalent ships without review, a foreign adversary is one exfiltration away from weaponizing it.

2. The AI czar changed hands

David Sacks, who championed the deregulatory stance, departed as White House AI czar in March. His replacements in practice — Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent — are both more comfortable with federal-review machinery than Sacks was. The departure matters more than the arrival; it removed a senior voice who would have killed a pre-release-vetting draft in committee.

3. Voters and Congress

The political backdrop is polling that shows a majority of registered voters believe AI's risks now outweigh its benefits — a reversal from 2024-25 numbers. Both parties in Congress have shown bipartisan appetite for AI oversight since the January job-displacement data. The executive branch can get ahead of that legislative pressure with a narrower, controllable working-group structure.

What the working group would actually review

The public reporting doesn't spell out the testing rubric, but the pattern of agencies named (NSA, NCO, ODNI) and the Mythos-as-catalyst framing suggest a specific shape. Here's the most likely evaluation scope based on what the involved agencies already assess elsewhere.

Review dimension	Lead agency	What "fail" would look like
Cyber offense uplift	NSA, White House National Cyber Office	Model autonomously finds zero-days at Mythos-scale or writes weaponizable exploit code
Bio/chem uplift	ODNI, WMD Center	Model provides meaningful synthesis uplift for dangerous pathogens or chemical weapons
Influence operations	ODNI, Foreign Malign Influence Center	Model can run at-scale persona networks or produce undetectable synthetic media
Autonomous agent capability	NSA, DoD CDAO	Model runs multi-day autonomous workflows against live infrastructure without human approval
Jailbreak resilience	Working group cross-agency	Universal jailbreak (see UK AISI's April 2026 finding) survives RLHF and Constitutional AI training
Export-control surface	Commerce BIS, Treasury OFAC	Model weights or distilled variants are exfiltratable by adversary state actors
Critical-infrastructure risk	CISA, NSA	Model agents can plan attacks against water, power, or financial-system infrastructure without refusal

The UK template, translated for DC

The UK's AI Safety Institute (renamed AI Security Institute in late 2025) has been running a soft version of pre-release evaluation since 2024, with voluntary submission from Anthropic, OpenAI, and Google DeepMind. Its universal-jailbreak finding in April 2026 — the one that broke every tested frontier model — is probably the most cited piece of public evidence that pre-release testing can find things the labs miss. A US executive order would likely codify the UK's voluntary-access pattern into something mandatory, at least for models distributed to federal buyers.

That creates an interesting asymmetry. Companies with mature internal evaluation stacks — Anthropic's evals team, OpenAI's Preparedness framework, Google DeepMind's Frontier Safety Framework — can plug into federal review with modest added friction. Companies without that infrastructure — Meta's Superintelligence Labs, xAI, the Chinese open-weight players — face a larger uplift cost or are simply locked out of federal procurement.

What the three companies are saying (and not saying)

The official reaction from Anthropic, Google, and OpenAI has been notably restrained. No public objections. No public endorsements. A White House official described the executive-order talk to reporters as "speculation" until the President announces — which reads as the companies and the administration agreeing to not pre-negotiate in public.

Privately, the incentive structure splits. Anthropic has spent two years pitching responsible-release discipline as a differentiator and has the best-developed evals program — formal federal review legitimizes that investment and handicaps faster-shipping competitors. OpenAI is in a harder position: the TAC program's existence suggests Altman is already accepting gated distribution as a new normal, and mandatory federal review extends that logic. Google DeepMind falls in the middle — strong internal evals, but a product-integration strategy (Gemini everywhere) that chafes against pre-release delay.

The DoD parallel track

The executive order isn't the only game. The Department of Defense is separately considering a policy that would make safety testing mandatory for any AI model distributed to federal or state government buyers. That track can move independently through normal DoD contract vehicles without requiring a White House signature. If the executive order stalls, the DoD rule achieves roughly 70% of the same effect for national-security buyers — the highest-value slice.

Combined with the May 1 Pentagon selection of eight firms for classified-network AI work (Anthropic excluded, which the White House has since softened with a six-month Claude-removal timeline and a Mythos carve-out), the federal buyer landscape for 2026 is visibly restructuring around risk review rather than raw capability.

Work with frontier AI, safely

Happycapy gives you Claude + GPT + Gemini in one workspace

$17 Pro · $167 Max · no per-seat lock-in

Try Happycapy →

What changes for the 2026 release calendar

If the executive order ships in the next 60 days — the working timeline implied by White House meeting cadence — three things shift.

Release-to-review lag. Frontier models intended for public release get a federal review window inserted before launch. That window is 30-90 days in the UK's voluntary version. US mandatory version probably starts at 14-30 days for covered models.
Model-variant proliferation. Labs ship the vetted public version + unvetted enterprise/federal variants + existing gated cyber/research variants. Distribution fragments further. TAC-style gating spreads.
China-gap narrative intensifies. Every week the working group takes, DeepSeek V4 and Kimi K2.6 ship further updates without review. Expect relentless pressure from the labs to accelerate the review cadence. See our piece on the Chinese open-weight wave.

The bigger story — from deregulation to gated access in one cycle

Rewind eighteen months. January 2025: the Trump administration rescinded the Biden AI executive order on day one and framed US AI leadership as deregulatory. David Sacks publicly mocked "safety theater." The dominant narrative was speed-to-market.

May 2026: the same administration is drafting pre-release vetting, the Pentagon is selectively excluding safety-focused vendors, OpenAI is running gated TAC distribution, Anthropic has Glasswing, the White House reopens talks with Anthropic because of Mythos, and the UK's voluntary-access pattern is being cited as the template.

Whatever happens next with the executive order, the direction of travel is clear: gated access and federal review are not fringe ideas anymore. They are the 2026 default.

FAQ

Is this executive order signed yet?

No. As of May 5, 2026, the order is in deliberation. A White House official described reporting about it as "speculation" until the President formally announces. The NYT story, sourced to US officials and people briefed on the talks, describes a framework under active discussion with tech executives.

Would open-weight models be covered?

The reporting doesn't specify. A working group modeled on the UK AISI approach would focus on frontier closed-weight models by default — the ones with capability claims high enough to trigger national-security review. Open-weight models published overseas (DeepSeek, Kimi, Qwen) effectively route around any US-only regime regardless.

How does this affect consumer AI tools?

For end-users paying for ChatGPT Plus, Claude Pro, or Gemini Advanced in 2026, the most likely visible effect is a small delay between a model's lab announcement and its consumer availability. Some cyber-adjacent capabilities may ship in gated enterprise variants before reaching consumer tiers, similar to how TAC already works for GPT-5.5 Cyber.