Summary
Forbes summary of a New York Times report (May 4, 2026) that the Trump administration is considering an executive order establishing a working group on AI and a formal government review process for new AI models before public release. The proposed regime is reportedly modeled on a U.K. multi-agency safety review. White House officials have held talks with Anthropic, Google, and OpenAI. The shift represents a reversal — Trump’s prior public posture (July 2025: “we have to grow that baby and let that baby thrive… we can’t stop it with foolish rules”) was deregulatory. The reversal is attributed in the reporting to the post-February breakdown between the Trump administration and Anthropic over Pentagon access to models, plus White House concern about “AI-enabled cyberattack” risk specifically tied to Anthropic’s Claude Mythos model.
Key Points
- Proposed executive order: would establish an AI working group of tech executives + government officials and a formal pre-release review process for new models
- Model: U.K. multi-agency AI safety review (UK Commons Library briefing CBP-10003 cited)
- Talks held: Anthropic, Google, OpenAI executives have met with White House officials on the plan
- Reversal context: Trump’s July 2025 posture was hands-off — “we have to grow that baby and let that baby thrive”; “we can’t stop it with foolish rules”
- The Anthropic break, February 2026: Anthropic refused a Trump administration request for “unrestricted access” to its AI models; Defense Secretary Pete Hegseth designated Anthropic a “supply chain risk to national security”
- Litigation: Anthropic filed suit to block the Pentagon designation; a federal appeals court refused (earlier in May) to stay the designation while the lawsuit proceeds
- Claude Mythos as the trigger: Despite the dispute, Anthropic cofounder Jack Clark said the company is in talks with the federal government about Claude Mythos — a model Anthropic claims could exploit security flaws in “every major operating system and web browser” and warned was too dangerous for public use
- March 2026 AI advisory panel: Trump appointed 13 members including Meta’s Mark Zuckerberg, Oracle’s Larry Ellison, Nvidia’s Jensen Huang, Dell founder Michael Dell — to advise on “science, technology, education and innovation policy”
- Legislative push: parallel administration framework calls on Congress to create a uniform national AI policy that would “supersede state rules” — i.e., federal preemption of state AI laws
- Pentagon AI vendor agreements (last week): OpenAI, Alphabet, Nvidia, SpaceX, Microsoft, Amazon, Reflection — all signed for “lawful operational use” of AI on classified DoD networks for “data analysis,” “situational understanding,” and “warfighter decision-making”
Newsletter Angles
- The reversal is doing two things at once: (1) creating a pre-release review chokepoint that the executive branch controls, and (2) preempting state-level AI laws via the parallel legislative framework. Both moves consolidate AI governance into a single point of executive leverage. The claimed safety justification (Claude Mythos cyber-exploit risk) is real, but the structure produced is political — it makes future model releases conditional on Trump-administration approval.
- Anthropic is the test case: The Pentagon “supply chain risk” designation against Anthropic is the first known case of an AI company being subjected to national-security suppliers’ restrictions for refusing a White House access demand. The pre-release review EO is the post-hoc generalization of that posture. This is the The Jawboning Papers / Federal Power as Political Instrument pattern applied to AI.
- Pair with the Pentagon vendor list: OpenAI, Alphabet, Nvidia, SpaceX, Microsoft, Amazon, Reflection — not Anthropic — signed the DoD classified-network agreements last week. The pre-release review framework is the carrot/stick complement: the firms that comply get DoD contracts; the firm that refused unrestricted access gets a supply-chain-risk designation. The picture is industrial policy by access-conditioning, not safety regulation in the conventional sense.
- The “U.K. model” comparison is doing rhetorical work: U.K. AI safety reviews are run by independent agencies under statutory authority, not by an executive working group of tech CEOs and political appointees. Citing the U.K. process while building a White House-controlled review is a category-shift that any newsletter coverage should call out.
- Federal preemption of state AI law is the deeper move: California’s SB 1047 era and various state AI bills are the direct target. Worth a piece pairing this with The TikTok Ban Is Just the Beta Test / The Anti-CBDC Surveillance State Act — federal-preemption-as-leverage is a recurring administration tactic.
- Claude Mythos is now the canonical “dangerous-model” reference: The administration is using Claude Mythos’s cyber-exploit capability claim to justify a broad regulatory regime affecting all models — not just frontier systems with verified offensive capabilities. The verification gap matters: Anthropic self-assessed Mythos as too dangerous; there’s no independent evaluation in the public record.
Entities Mentioned
- Donald Trump — administration policy lead
- Anthropic — the company that refused unrestricted access
- Jack Clark — Anthropic cofounder; named as in talks about Mythos
- Pete Hegseth — Defense Secretary; designated Anthropic supply-chain risk
- OpenAI — White House meetings; DoD vendor agreement
- Google / Alphabet — White House meetings; DoD vendor agreement; deferred stub if not yet present
- Meta / Mark Zuckerberg — AI advisory panel member
- Oracle / Larry Ellison — AI advisory panel member; deferred stub
- Nvidia / Jensen Huang — AI advisory panel; DoD vendor; deferred stub
- Dell / Michael Dell — AI advisory panel; deferred stub
- Microsoft — DoD vendor agreement
- SpaceX — DoD vendor agreement; deferred stub if not yet present
- Amazon — DoD vendor agreement (existing in wiki context for CARES Act)
- Reflection — DoD vendor agreement; deferred stub (likely Reflection AI)
Concepts Mentioned
- Claude Mythos — the named “dangerous model” used to justify the regime (entity, not concept; existing wiki page)
- Frontier AI — broader concept the regulation targets
- AI executive order / pre-release review — deferred concept (nascent — wait for the actual EO text)
- Federal preemption of state AI law — deferred concept; pair with telecom / surveillance preemption history
- Faith and Militarism — adjacent (DoD-AI integration is the operational context)
Quotes
“The Trump administration is considering an executive order to establish a working group on AI featuring tech executives and government officials, including plans for a formal government review process for new AI models.” — NYT via Forbes
“We have to grow that baby and let that baby thrive… We can’t stop it with foolish rules and even stupid rules.” — Trump, July 2025 (the prior posture)
“[Any rules] would have to be ‘more brilliant than even the technology itself.‘” — Trump, July 2025
Anthropic / Jack Clark on Claude Mythos: the model “could exploit security flaws in ‘every major operating system and web browser’ and warned it was too dangerous for public use”
Notes
- Source tier: Forbes is the summary; the underlying NYT report (Cecilia Kang/Tripp Mickle/etc., NYT 5/4/2026, “White House Considers Vetting AI Models Before They Are Released”) is the primary. Worth ingesting NYT directly if writing on this — Forbes cites U.S. officials and “people briefed on the talks” sourced through NYT.
- Verification status: An EO has not been issued — this is a “considering” report. The Claude Mythos capability claim is Anthropic’s self-assessment, not independently verified.
- Companion sources:
- Anthropic entity page — captures the company’s lineage
- Claude Mythos (existing wiki source) — captures the model and the cyber-exploit claim
- What’s missing from the public record:
- The actual EO text or draft language
- Whether the proposed review covers open-weights models, only frontier/closed models, or all
- The legal basis for federal preemption of state AI law (the framework is “calling on Congress” but the EO route would be limited to federal contractors / national-security review)
- Whether the AI advisory panel has formal regulatory authority or is purely advisory
- Track: NYT follow-ups; any actual EO publication; Anthropic v. DoD lawsuit docket for the supply-chain designation challenge.