> ## Content Index
> Fetch the complete content index at: https://ppc.land/llms.txt
> Use this file to discover other available public pages before exploring further.

# Trump signs AI order reviving the safety review he abolished 17 months ago
- URL: https://ppc.land/trump-signs-ai-order-reviving-the-safety-review-he-abolished-17-months-ago/
- Published: 2026-06-03T05:48:21.000Z
- Updated: 2026-06-03T05:48:21.000Z
- Description: Trump signs AI order with voluntary 30-day frontier model review as Anthropic's Mythos alarmed Washington and prompted EU access talks with the US government.
- Author: Luis Rijo
- Tags: AI

*President Donald Trump on June 2 signed a new executive order on artificial intelligence, establishing a voluntary framework under which AI companies can submit their most powerful models to the federal government for review up to 30 days before public release - a structural approach that closely mirrors the Biden-era mandatory requirement his administration revoked on January 20, 2025\. The signing came one day after the European Commission confirmed it had reached an agreement with Anthropic to access the same Mythos model that, since its April 2026 launch, has driven concern in both Washington and Brussels.*

The order, titled [Promoting Advanced Artificial Intelligence Innovation and Security](https://www.whitehouse.gov/presidential-actions/2026/06/promoting-advanced-artificial-intelligence-innovation-and-security/?ref=ppc.land), was signed on June 2, 2026, at the White House, without a livestream or public ceremony. It combines a **cybersecurity** clearinghouse, new directives for federal agency defenses, an AI-enabled hiring expansion, and the headline mechanism: the voluntary prerelease review framework for what the order designates as **covered frontier models**.

## From Biden's mandate to a voluntary framework

The gap between where AI governance policy started under Trump and where it landed on June 2 amounts to more than a reversal. It is a full circle, completed in 17 months.

On January 20, 2025, Trump's first-day executive orders included the revocation of Biden's AI executive order, which had required developers of powerful AI systems to conduct safety tests and share results with the federal government before public deployment. The Trump administration characterized those requirements as "unnecessarily burdensome" and said they would stifle private sector innovation. According to the administration's own Day 1 language, the Biden rules imposed "onerous and unnecessary government control over the development of AI."

The order creates a voluntary version of almost exactly what was abolished. The mechanism is narrower - 30 days instead of the Biden-era mandatory approach, voluntary instead of compulsory, and without the civil rights and algorithmic discrimination provisions that the Biden order contained. But the underlying logic is structurally identical: the government should have visibility into powerful AI models before they reach the public.

A previous draft of the order, obtained and published by Politico on May 22, 2026, had proposed a 90-day review window. According to Craig McDonogh, founder of Ethicore Advisors, Trump had been slated to sign that more demanding version in late May but delayed after industry pushback, including from venture capitalist and former White House AI advisor David Sacks. The signed order cuts the review period to 30 days. As [PPC Land reported when the May 22 draft was shelved](https://ppc.land/trumps-ai-cybersecurity-order-what-the-shelved-draft-actually-says/), the draft had proposed the 90-day window for access to covered frontier models and outlined the same AI cybersecurity clearinghouse structure that appears in the final text.

## What the order actually directs

The text sets out a series of agency mandates across five sections, nearly all organized around either cyberdefense upgrades or the new **frontier model** framework.

Within 30 days of signature - meaning by approximately July 2, 2026 - multiple agencies face immediate compliance deadlines. According to the order, the Secretary of Homeland Security, through the Director of the Cybersecurity and Infrastructure Security Agency (CISA), must release **Binding Operational Directives** to expedite the cyber defense of civilian federal government information systems, establish or expand federal programs for AI-enabled defensive tools, and facilitate access to cybersecurity tools for federal agencies, state and local authorities, and operators of critical infrastructure. The order specifically names rural hospitals, community banks, and local utilities as examples of critical infrastructure operators that should gain access to these tools and services.

Also within 30 days, the Secretary of the Treasury, in consultation with the National Cyber Director and the directors of the NSA and CISA, must form an AI **cybersecurity clearinghouse**. According to the order, this clearinghouse will coordinate and deconflict scanning for software vulnerabilities, discover and validate those vulnerabilities, and coordinate and prioritize remediation and distribution of patches - all in voluntary collaboration with the AI industry and operators of critical infrastructure.

Within 60 days - approximately August 1, 2026 - the Secretary of the Treasury, working with the Director of the NSA and the Director of CISA, must develop and maintain a **classified benchmarking process** to assess the advanced cyber capabilities of AI models. That process will determine the threshold at which a model is designated a covered frontier model. The determination of which models meet that designation will be made by the Director of the NSA, in consultation with the National Cyber Director, the APST, and the Director of CISA.

Within the same 60-day window, the same agencies must design the voluntary framework through which AI developers can engage the federal government to determine whether a model under development meets the covered frontier model designation. Under this framework, developers would provide the federal government with access to covered frontier models - subject to confidentiality, cybersecurity, insider-risk, and intellectual property protections - for a period of up to 30 days before the planned release to other trusted partners. Developers and the government would then collaborate to select trusted partners who will have early access to covered frontier models.

The order is explicit on one point: nothing in the section on secure frontier model deployment shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement for the development, publication, release, or distribution of new AI models.

Also within 60 days, the Director of the Office of Personnel Management must expand the United States Tech Force Information Cybersecurity Specialist hiring and placement pathways.

## The EU sought access - and needed US permission

The day before the White House signed this order, a parallel diplomatic process concluded in Brussels. The European Commission [confirmed to CNBC](https://www.cnbc.com/2026/06/01/anthropic-eu-ai-mythos-access-advanced-model.html?ref=ppc.land) on June 1, 2026, that it had "several productive meetings" with Anthropic and reached an agreement for EU access to the Mythos model.

According to Thomas Regnier, the EU's tech sovereignty spokesperson, the commission aims to get a clearer idea of the potential risks the technology poses. "Let's not forget that Mythos is not one off, a new wave of powerful models are coming to the market," Regnier said. "This is a shared challenge, and we are intensifying our discussions with like-minded partners, including the United States."

The process was not straightforward. According to people familiar with the matter cited by CNBC, Anthropic told the European Commission it needed US government permission to share the model. The US government, those people said, is against sharing the powerful model with non-US governments broadly, as it aims to remain the dominant AI leader. The EU had to seek permission from the US administration in order to gain access. Official terms and conditions of the agreement were not fully clear as of June 1.

The EU's situation with Mythos contrasts with its earlier experience with OpenAI's GPT-5.5-Cyber model, to which the commission had already been granted access in May 2026\. According to CNBC's reporting, the commission had four or five meetings with Anthropic at the time of that comparison, and the Anthropic discussions were at a "different stage."

Governments, banks, and technology firms were alarmed by Mythos when it was first released in April 2026 as part of Anthropic's **Project Glasswing** program - a limited partner initiative for defensive security tasks. The model is capable of uncovering thousands of previously unknown software vulnerabilities, raising concerns about misuse by bad actors who could accelerate cybercrime at scale.

## The Anthropic factor in US policy

According to Matthew Ferren, an international affairs fellow in national security at the Council on Foreign Relations, the order reflects an administration trying to sustain its deregulatory, innovation-first posture while confronting the novel cyber risks posed by powerful new tools like Anthropic's Claude Mythos Preview.

The role of Mythos in prompting this policy shift is direct. In April 2026, according to McDonogh, Mythos sent shockwaves through Washington with its demonstrated ability to find critical and severe vulnerabilities in widely used operating systems. [According to CFR experts](https://www.cfr.org/articles/assessing-trumps-executive-order-on-ai-oversight?ref=ppc.land), concerns about Mythos's capability to autonomously identify and exploit hidden vulnerabilities in real-world software were among the factors that directly accelerated the order's development. A single AI system powerful enough to find thousands of critical software vulnerabilities changed the political calculus for an administration that had publicly committed to keeping government out of AI development.

The government-Anthropic relationship by that point had become extraordinarily complicated. As PPC Land reported in detail, [the Pentagon in late February 2026 designated Anthropic a supply-chain risk to national security](https://ppc.land/anthropic-sues-us-government-over-claude-ban-and-supply-chain-label/) after the company refused to remove two restrictions from its usage policy - one prohibiting use for mass surveillance of Americans and one prohibiting use in fully autonomous lethal weapons. That dispute led Anthropic to sue the US government. A [federal court subsequently blocked the Pentagon's blacklisting of Anthropic](https://ppc.land/judge-blocks-pentagon-from-blacklisting-anthropic-over-ai-safety-stance/), ruling the action was First Amendment retaliation. Officials have credited Palantir's Maven system, which incorporates Anthropic's Claude, with compressing military targeting cycles in Iran from days to minutes. Meanwhile, the Pentagon had been seeking nearly $30 billion to build its own AI infrastructure.

The order, signed without fanfare, appears to represent a partial normalization of the government-Anthropic relationship. Anthropic, which already has a working relationship with the NSA and military through these arrangements, has an obvious institutional incentive to participate in a review process that validates its safety posture and differentiates it from competitors.

## The voluntariness question

The word "voluntary" appears in the order repeatedly, and its meaning in practice depends heavily on the incentive structures surrounding it.

According to Vinh Nguyen, a senior fellow for AI at the Council on Foreign Relations, the voluntary framework - where developers grant the government secure access to covered models up to 30 days before their release to trusted partners - is the mechanism that converts opacity into observability. Whether it works depends less on the legal text than on whether both sides treat that access as genuine collaboration.

Ferren's [CFR analysis](https://www.cfr.org/articles/assessing-trumps-executive-order-on-ai-oversight?ref=ppc.land) notes that US frontier labs will likely participate in the testing regime voluntarily, if only to forestall more invasive regulation later. But other models may soon replicate the cyber capabilities that prompted this order. Google's threat intelligence team has documented state-aligned actors already using frontier models to automate cyberattacks, and researchers have shown that Mythos-style vulnerability reasoning can be reproduced with open-weight systems.

McDonogh's analysis makes a related observation about the compliance landscape. A voluntary review framework that primarily targets responsible actors while leaving less responsible ones unaffected shifts competitive dynamics while providing limited actual safety benefits. Companies already doing extensive prerelease safety evaluation will participate and receive government endorsement. Companies doing less will decline and face no consequences for doing so. The companies least likely to participate are those releasing smaller models without the reputational infrastructure that makes government endorsement valuable - and frontier labs outside the United States, particularly in China, which will not submit their models to the US government for review under any circumstances.

## NSA and Treasury: an unusual pairing

One structural choice in the order that has drawn analyst attention is the prominent operational role assigned to the Treasury Department. According to Ferren's CFR analysis, Treasury is assigned a leading role in coordinating the AI cybersecurity clearinghouse and the frontier model testing framework, with a prominence relative to more obvious parties like CISA or the Office of the National Cyber Director. That may reflect, Ferren suggests, that Treasury is one of the few places in the federal government where institutional cybersecurity capacity remains intact after significant workforce reductions over the past 18 months.

The classified benchmarking process will be run by the NSA, with the Director of NSA making the final determination on which models qualify as covered frontier models. That places the intelligence community at the center of the review mechanism, which has implications for how AI companies will approach participation. The order's access provisions include requirements for confidentiality, insider-risk protections, and intellectual property nondisclosure - provisions evidently designed to address industry concern that government review creates exposure for proprietary model architectures.

## Context: a year and a half of executive actions

The June 2 order does not arrive in isolation. It follows a sequence of executive actions on AI that the Trump administration has taken since January 2025.

According to the White House fact sheet released alongside the order, the administration eliminated Biden's AI policies immediately upon taking office in January 2025\. In July 2025, Trump released the AI Action Plan, which called for examining and removing regulations that hinder American leadership in AI technology. Also in July 2025, he signed an executive order preventing the federal government from using AI models that include ideological biases or social agendas. In December 2025, he signed an executive order to protect American AI innovation from an inconsistent and costly compliance regime resulting from varying state laws.

The White House also notes that in March 2026, Trump released his National Cyber Strategy for America, calling for unprecedented coordination across government and the private sector on cybersecurity investment. That same month, the administration unveiled a comprehensive national legislative framework addressing pressing AI policy topics - a document that [PPC Land covered in March 2026](https://ppc.land/white-house-ai-framework-targets-state-laws-child-safety-and-copyright/) as addressing child safety, intellectual property, free speech, workforce development, and federal preemption of state-level AI regulations.

The Trump administration also announced agreements in May 2026 with Google DeepMind, Microsoft, and Elon Musk's xAI that will allow the government to assess artificial intelligence models before they are released publicly, according to CNBC. The signed order formalizes the framework under which such assessments will operate.

The executive order is the most operationally specific AI action the administration has taken. Where previous orders directed agencies to act, this one creates specific timelines and measurable deliverables: binding directives within 30 days, a clearinghouse formed within 30 days, classified benchmarks and a voluntary framework designed within 60 days, cybersecurity hiring pathways expanded within 60 days.

## What remains unresolved

According to [Nguyen's CFR analysis](https://www.cfr.org/articles/assessing-trumps-executive-order-on-ai-oversight?ref=ppc.land), the most difficult problem the order asks agencies to solve is defining what counts as a covered frontier model. Frontier AI systems are probabilistic, goal-directed, increasingly autonomous, and opaque. They do not have fixed capability ceilings and exhibit emergent behaviors that shift with scale, fine-tuning, and deployment context. A model that appears unremarkable in isolated testing could become a potent cyber tool when integrated into an autonomous pipeline with access to real-world digital infrastructure.

Nguyen argues that the covered frontier model designation needs to be understood as covering not just a model in isolation but an AI system - including its integrated models, data pipelines, and deployment architecture - performing at the state of the art in domains relevant to national security. If the definition is too narrow, genuinely dangerous capabilities may ship without evaluation. If too broad, the evaluation process exhausts the limited talent available to do this work.

Frontier AI capabilities advance on a timeline measured in months, not years. The institutions charged with evaluation will need to match that tempo or they will assess yesterday's models against yesterday's threats.

Ferren's analysis adds a further structural concern: finding vulnerabilities is often the easy part. Consistent patching remains an unsolved problem, particularly for open-source projects and under-resourced critical infrastructure operators such as school districts and water treatment facilities. Many remain more vulnerable to basic cyber hygiene shortfalls than to any AI-discovered vulnerability. The federal government, which has cut its cybersecurity workforce substantially over the past 18 months, faces a significant organizational challenge in coordinating the nationwide software hardening campaign the order envisions.

## Implications for the marketing technology industry

The marketing technology industry is not named in the executive order, and the order's operative provisions target national security systems rather than commercial AI deployments. The direct compliance obligations on marketing organizations deploying AI tools are unchanged. The order does not address algorithmic bias, consumer transparency, employment discrimination, or the areas that state laws and EU regulations have been building frameworks around.

The indirect implications run deeper, however. The EU AI Act compliance framework, which [PPC Land has tracked through the Digital Omnibus discussions and the May 2026 agreement on high-risk deadlines](https://ppc.land/eu-ai-act-gets-its-first-real-haircut-high-risk-deadlines-pushed-to-2027/), remains on its own trajectory. The EU's decision to seek access to Mythos - and to need US government permission to do so - illustrates how the most powerful AI systems are now effectively treated as strategic assets that travel through diplomatic as well as commercial channels.

The voluntary standard set by the US order nonetheless creates a visible floor for what responsible AI deployment is expected to look like. Enterprise procurement teams, investors, and regulators who evaluate AI companies will note whether they participated in the government's review framework. According to McDonogh, voluntary today has a way of becoming expected, then standard, then mandatory - particularly if any model that skipped the review is subsequently associated with harm. Organizations building genuine AI governance infrastructure now, rather than waiting for a regulatory trigger, are positioning themselves for a compliance landscape that is moving toward mandatory oversight regardless of its current voluntary framing in the US.

## Timeline

- **January 20, 2025** \- Trump revokes Biden's AI executive order, which had mandated that developers conduct safety tests and share results with the federal government before public deployment, calling it "unnecessarily burdensome"
- **January 23, 2025** \- Trump signs Executive Order 14179, "Removing Barriers to American Leadership in Artificial Intelligence," directing the development of an AI action plan
- **July 23, 2025** \- Trump releases the AI Action Plan and signs three AI-related executive orders, including one on promoting the export of the American AI technology stack
- **December 11, 2025** \- Trump signs an executive order to protect American AI innovation from varying state law compliance regimes
- **February 16-27, 2026** \- Pentagon dispute with Anthropic escalates, leading to the designation of Anthropic as a supply-chain risk to national security; [PPC Land covers the lawsuit and judicial block in detail](https://ppc.land/anthropic-sues-us-government-over-claude-ban-and-supply-chain-label/)
- **March 2026** \- Trump releases the National Cyber Strategy for America and a comprehensive national legislative framework for AI; [PPC Land covers the White House AI framework](https://ppc.land/white-house-ai-framework-targets-state-laws-child-safety-and-copyright/)
- **March 28, 2026** \- Federal judge blocks the Pentagon's supply-chain risk designation against Anthropic in a First Amendment ruling; [reported by PPC Land](https://ppc.land/judge-blocks-pentagon-from-blacklisting-anthropic-over-ai-safety-stance/)
- **April 2026** \- Anthropic launches Mythos Preview through Project Glasswing, a limited cybersecurity partner program; the model draws immediate concern in Washington and internationally for its ability to find thousands of previously unknown software vulnerabilities
- **May 2026** \- The European Commission gains access to OpenAI's GPT-5.5-Cyber model; the Trump administration announces pre-release assessment agreements with Google DeepMind, Microsoft, and xAI
- **May 22, 2026** \- A draft executive order proposing a 90-day voluntary review window is shelved after industry pushback; [PPC Land reports on the shelved draft's contents](https://ppc.land/trumps-ai-cybersecurity-order-what-the-shelved-draft-actually-says/)
- **June 1, 2026** \- The European Commission confirms agreement with Anthropic for EU access to Mythos, following weeks of talks that required US government permission; EU tech sovereignty spokesperson Thomas Regnier calls Mythos "not one off" and describes it as "a shared challenge"
- **June 2, 2026** \- Trump signs "Promoting Advanced Artificial Intelligence Innovation and Security," establishing a 30-day voluntary review framework for covered frontier models, an AI cybersecurity clearinghouse led by Treasury, classified benchmarking run by the NSA, and agency cyber defense mandates with 30- and 60-day deadlines; [EU AI Act high-risk deadline framework as tracked by PPC Land](https://ppc.land/eu-ai-act-gets-its-first-real-haircut-high-risk-deadlines-pushed-to-2027/) remains on its own timeline

## Summary

**Who:** President Donald Trump, acting through the White House, with operational roles assigned to the Secretary of the Treasury, the Secretary of Homeland Security through CISA, the Director of the NSA, the National Cyber Director, the Director of OMB, and the Attorney General. AI companies, particularly frontier model developers such as Anthropic, OpenAI, Google DeepMind, Microsoft, and xAI, are the primary non-government actors addressed. The European Commission and its tech sovereignty office are a parallel actor, having secured access to Anthropic's Mythos model one day before the order's signing.

**What:** An executive order titled "Promoting Advanced Artificial Intelligence Innovation and Security," signed on June 2, 2026, establishing a voluntary 30-day prerelease review framework for designated covered frontier AI models, a new AI cybersecurity clearinghouse coordinated by Treasury, a classified benchmarking process run by the NSA, binding operational directives for federal civilian cybersecurity, and expanded federal cybersecurity hiring pathways. The EU's concurrent agreement for Mythos access underlines the international dimension of frontier model governance that the US order addresses domestically.

**When:** Signed on June 2, 2026\. Key deadlines embedded in the order are 30 days from signature for the cybersecurity clearinghouse formation, CISA binding directives, and National Security Systems cyber defense prioritization; 60 days from signature for the classified benchmarking process, the voluntary framework design, and the cybersecurity hiring expansion. The EU Mythos access agreement was confirmed on June 1, 2026.

**Where:** The order operates across all US federal government information systems, civilian federal agencies, state and local authorities, and operators of critical infrastructure. The voluntary frontier model review framework applies to AI developers operating in the United States. The order was signed at the White House without a public ceremony. The EU access discussions took place between the European Commission and Anthropic, with US government permission required as an intermediate step.

**Why:** The administration has cited mounting national security concerns about the advanced cyber capabilities of frontier AI models - specifically the ability of powerful models to autonomously identify and exploit software vulnerabilities in real-world systems. According to analysts at the Council on Foreign Relations, the shift was driven in part by Anthropic's Claude Mythos Preview demonstrating a level of cybersecurity capability that changed the political and policy calculus for an administration that had publicly committed to keeping government out of AI development. The EU's parallel effort to access Mythos, requiring US government intermediation, confirms that the most powerful AI models are now treated as strategically sensitive in ways that extend beyond domestic regulation.