Last reviewed
Jun 3, 2026
Sources
11 citations
Review status
Source-backed
Revision
v1 · 1,995 words
Add missing citations, update stale details, or suggest a clearer explanation.
Promoting Advanced Artificial Intelligence Innovation and Security is a United States presidential executive order signed by President Donald Trump on June 2, 2026. The order asks artificial intelligence developers to voluntarily hand the federal government early access to their most capable systems, called "covered frontier models," for up to 30 days of security testing before a wider release. It also directs agencies to build benchmarks for the cyber capabilities of frontier models, to stand up an "AI cybersecurity clearinghouse" for finding and patching software vulnerabilities, and to harden federal networks. The order explicitly rules out any mandatory licensing or preclearance regime, a notable retreat from a stricter draft that the White House had drawn up weeks earlier and then shelved under pressure from the AI industry.
The order arrived at an awkward moment for the administration's hands-off posture on AI. It followed months of alarm inside government over Anthropic's Claude Mythos model and its demonstrated ability to find and exploit software flaws on its own. The result is a document that tries to thread a needle: give defenders a head start on dangerous capabilities without imposing the kind of safety mandates the White House had spent the previous year fighting at the state level.
The order frames itself as a partnership rather than a regulation. Its stated policy is that "the United States continues to lead the world in Artificial Intelligence because of the enormous talent and innovation of our AI industry, and because we refuse to stifle this innovation with overly burdensome regulation." The administration pledges to "promote AI innovation and security by working collaboratively with the private sector to modernize government and private sector information systems and harden them against external threats," to protect American intellectual property from theft by adversaries, and to build out the country's AI-enabled defensive capabilities.[1][2]
In practice the order does three things. First, it sets up a voluntary channel for the government to review the most capable new models before they ship. Second, it tells national security and cybersecurity agencies to figure out which models are dangerous enough to warrant that review, using a classified benchmarking process focused on offensive cyber. Third, it pours effort into federal cyber defense, directing agencies to expand AI-enabled security tools, prioritize the protection of national security and civilian systems, and share threat information more aggressively.[1][3]
| Provision | Detail |
|---|---|
| Voluntary early access | Developers may give the government access to "covered frontier models" for up to 30 days before releasing them to other trusted partners, subject to confidentiality, cybersecurity, insider-risk, and intellectual property protections.[1][4] |
| Cyber-capability benchmarks | The NSA, with the Treasury Department and the Cybersecurity and Infrastructure Security Agency (CISA), develops a classified, multilayered benchmarking process to assess a model's advanced cyber capabilities and decide which systems count as covered frontier models.[3][4][5] |
| AI cybersecurity clearinghouse | Treasury, NSA, and CISA establish a clearinghouse, in voluntary collaboration with industry, to coordinate and deconflict scanning for software vulnerabilities, validate them, and prioritize the distribution of patches.[1][5] |
| Trusted partners program | Developers may work with agencies to select government-vetted "trusted partners" who get coordinated early access to covered models, intended to strengthen the cybersecurity of critical infrastructure.[3][5] |
| Federal cyber hardening | CISA issues binding operational directives; OMB, the National Cyber Director, and the Secretary of Homeland Security expand AI security tools for federal, state, local, and critical-infrastructure operators, with 30- and 60-day action deadlines.[4][5] |
| No mandatory licensing | "Nothing in this section 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, including frontier models."[1][6] |
The order leaves the most consequential question, what actually qualifies as a "covered frontier model," to later agency work. A legal analysis by the firm WilmerHale noted that the order does not define the term precisely and makes the criteria "a focal point of agency rulemaking" in the months ahead, with the NSA director leading the determination of which systems get scrutiny and which trusted partners receive early access.[4]
The defining feature of the final order is what it does not do. It does not require any company to submit a model, does not gate a launch on government approval, and does not create a licensing scheme. The White House underscored the point in a social media post the day of the signing, saying the government is "NOT conducting oversight of all new models, as that level of government overreach would have chilling effects on free speech."[6]
This was a deliberate climbdown. An earlier draft would have given the government up to 90 days to review advanced models before release, a window the industry warned was long enough to slow product cycles and hand a competitive edge to rivals abroad. The final text cut that to 30 days and stripped out the harder mandates. Trump had been expected to sign a version of the order on May 21 at an Oval Office event with tech executives, but he canceled the ceremony, telling reporters he did not support "certain aspects of it." He tied the reversal to the global race with China: "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."[6][7] Twelve days later he signed the narrowed version in private, without the planned ceremony.[6]
The order is hard to understand without the Claude Mythos episode that preceded it. Anthropic previewed Mythos in April 2026 and showed it autonomously finding and exploiting vulnerabilities, including decades-old flaws in widely used software, with little human input.[8] The demonstration landed as a live national security concern rather than a research curiosity. Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell convened an urgent meeting with Wall Street chief executives to warn them about the risk that AI could surface vulnerabilities across the financial sector's software faster than defenders could patch them.[8][9]
The worry was not hypothetical for the government either. Anthropic held cyberthreat briefings with federal agency chief information officers on May 7 and 8, 2026, walking them through how to defend against threats powered by advanced models including the Mythos preview, and the House Homeland Security Committee received a similar briefing.[9] The company had earlier disclosed GTG-1002, a state-linked espionage campaign in which an attacker used its models to automate much of an intrusion operation, which sharpened fears about offensive AI in cybersecurity.
That backdrop also explains a tension running underneath the order. The Pentagon had labeled Anthropic a supply-chain risk and barred defense contractors from using its technology, and Anthropic sued the administration to reverse the designation, with the litigation still ongoing as the order was signed.[8] Critics noted the irony of an order that leans on a voluntary review process whose value depends on cooperation from a company the Department of War had tried to freeze out.
The order is the latest entry in a year and a half of Trump-era AI policy that has mostly run in the opposite direction, toward deregulation and away from oversight. In July 2025 the administration released America's AI Action Plan, built around accelerating innovation, building infrastructure, and exporting American AI, alongside an executive order barring the federal government from procuring models with "ideological biases or social agendas." In December 2025 Trump signed Executive Order 14365, "Ensuring a National Policy Framework for Artificial Intelligence," which sought to preempt the patchwork of state AI laws by setting up an AI Litigation Task Force to challenge them and threatening to withhold federal broadband funds from states with rules the administration disliked.[10] That order treated state-level regulation as the main obstacle to AI leadership.
Set against that record, the June 2026 order is a partial reversal of tone. The White House fact sheet still drew a contrast with its predecessor, saying that "unlike the Biden Administration's top-down regulatory approach, President Trump is working hand-in-hand with American industry to strike the right balance between innovation and national security."[2] But the order represents the administration's first move toward any federal review of frontier models, even a voluntary one, after eighteen months spent dismantling the previous government's guardrails.
Response split along familiar lines. Industry largely welcomed the light touch. OpenAI's chief global affairs officer Chris Lehane called the order "an important step" and argued that AI governance frameworks should be "developed through democratic institutions, informed by technical expertise."[6]
Several lawmakers and advocacy groups said voluntary was not enough. Senator Josh Hawley (R-Mo.) backed the order but said he would press to make the reporting mandatory through legislation.[3] Senator Mark Warner (D-Va.), of the Senate Intelligence Committee, welcomed the policy while criticizing the administration for having "belatedly discovered" the case for AI oversight after spending a year tearing down Biden-era rules.[6] Brendan Steinhauser of the Alliance for Secure AI said "Congress must now codify" the approach with legislation, and Brad Carson of Americans for Responsible Innovation urged lawmakers to "make these protections mandatory."[3]
Others worried less about the order being too weak and more about how its discretion could be used. Juan Londoño of the Cato Institute called it "a step in the right direction" but warned that handing the NSA director authority to decide which models face review sets a "dangerous precedent" that a future administration could "weaponize" against disfavored companies such as Anthropic.[6]
Analysts at the Council on Foreign Relations framed the order as a bet on timing. Matthew Ferren described it as "an attempt to engineer a cybersecurity window of opportunity," giving defenders early access to dangerous models before attackers catch up, while cautioning that the voluntary framework may not hold once "other models replicate their cyber capabilities" and that "consistent patching remains an unsolved problem, particularly for open-source projects." His colleague Vinh Nguyen pointed to the deeper difficulty of defining a covered frontier model at all, given that modern systems are "probabilistic, goal-directed, increasingly autonomous" and have no "fixed capability ceilings."[11]
The order matters less for what it mandates, which is almost nothing, than for what it signals. After staking its AI agenda on the premise that regulation is the threat and speed is the asset, the administration conceded that the most capable models can pose national security risks serious enough to warrant a federal look before release. Whether a voluntary 30-day window is enough to manage those risks, or whether it becomes the floor for the mandatory regime several senators are already drafting, will depend on how agencies define a covered frontier model and how many developers choose to take part.