Trump’s New AI Order Gives NSA Access, Skips Mandatory Licensing

TL;DR

• Trump signed an executive order creating a voluntary federal review framework for advanced “covered frontier” AI models with cybersecurity risks—no mandatory licensing required.
• The order directs agencies to strengthen federal cyber defenses within 30 days and gives NSA and Treasury central oversight roles in assessing which models qualify for review.
• Major U.S. labs like OpenAI, Anthropic, and Google will face expectations for pre-release red-teaming on cyber capabilities, with developers potentially granting government access for up to 30 days.
• Critics will argue the voluntary approach is too weak, while civil libertarians worry about expanded NSA access to cutting-edge models.

Trump’s Frontier Model Framework Splits the Difference

The White House issued an executive order titled “Promoting Advanced Artificial Intelligence Innovation and Security” that establishes the first formal federal process for reviewing advanced AI models before they ship—but keeps it voluntary. The order targets so-called “covered frontier models” that pose cybersecurity risks, requiring developers to coordinate with U.S. security agencies including the NSA and Treasury without triggering the mandatory licensing regime that industry has fought against for two years. Agencies have 60 days to develop a classified benchmarking process for designating which models qualify as covered frontier systems.

The order directs agencies to strengthen federal cyber defenses within 30 days, creates an AI cybersecurity clearinghouse, and prioritizes enforcement of existing criminal statutes against AI-enabled cyberattacks. Developers may provide the government pre-release access to covered models for cybersecurity testing for up to 30 days under the voluntary framework. It’s the most significant U.S. federal AI governance move since the Biden administration secured voluntary commitments from major labs in 2023.

Why the NSA and Treasury Just Became Gatekeepers

Here’s where it gets interesting. The order explicitly elevates the National Security Agency and Treasury Department into central oversight roles—agencies that historically haven’t played lead parts in AI policy. That’s a deliberate signal that the Trump administration views frontier models primarily through a national security and economic competitiveness lens, not a consumer protection or civil rights frame. The NSA brings signals intelligence and offensive cyber expertise; Treasury brings sanctions enforcement and financial system defense.

The 60-day timeline to develop classified benchmarking criteria means the definition of “covered frontier model” will be hashed out behind closed doors by security officials, not in public rulemakings. And that’s going to make transparency advocates nervous. Who decides whether GPT-5 or Claude 4 crosses the threshold? What compute levels trigger review? The order gives agencies broad discretion to define those red lines without publishing the methodology.

But—and this matters—the framework is voluntary. No lab is legally compelled to submit models for pre-release review. The order explicitly rejects mandatory AI licensing or preclearance, which tells you everything about the political calculus here. The administration wants visibility into the most dangerous capabilities without handing ammunition to critics who’d call it a federal chokehold on innovation.

What This Means for OpenAI, Anthropic, and the Rest

Major U.S. AI labs are the obvious targets for this voluntary program. OpenAI, Anthropic, Google DeepMind, Meta, and any startup training models north of 10^26 FLOPs will face pressure—informal but real—to participate in pre-release cybersecurity testing. The order doesn’t name compute thresholds, but the 60-day classified benchmarking process will almost certainly land somewhere in that range based on international precedent from the U.K. and U.S. AI safety summits.

The competitive stakes are straightforward. Labs that cooperate signal responsibility and buy goodwill with federal agencies that control cloud contracts, research grants, and—if things go sideways—investigative scrutiny. Labs that don’t cooperate risk looking reckless if a model they ship later gets caught enabling a ransomware campaign or critical infrastructure attack. It’s soft power, but it’s power.

The order also positions the U.S. as pursuing a lighter-touch, security-focused approach compared with the EU’s AI Act compliance regime. European rules impose mandatory conformity assessments, transparency obligations, and potential fines up to 6% of global revenue for high-risk systems. This executive order asks nicely and threatens nothing—at least not explicitly. That regulatory gap could influence where companies choose to launch or test their most powerful systems, especially if Brussels starts enforcing its rules aggressively.

I think the voluntary framing is a feature, not a bug. The administration knows it can’t pass AI legislation through Congress right now, so it’s building a de facto review process using executive authority and industry peer pressure. It’s the regulatory equivalent of a handshake deal—effective until someone defects.

The Licensing Fight That Didn’t Happen

Previous U.S. AI actions largely relied on voluntary commitments secured by the Biden administration and sectoral guidance, with no binding federal AI law. Internationally, debates over frontier model licensing and compute thresholds have intensified, with the U.K. and U.S. AI safety summits highlighting cyber risks from advanced models. This executive order builds on those concerns by formalizing a federal framework to assess high-risk models’ cybersecurity implications without creating a full licensing system.

The explicit rejection of mandatory licensing is the dog that didn’t bark. For two years, AI safety advocates and some lawmakers have pushed for a federal preclearance regime modeled on drug approval or financial regulation—submit your model, prove it’s safe, get a green light to deploy. Industry fought that tooth and nail, arguing it would freeze innovation and hand regulatory capture to incumbents. This order sides with industry on structure while giving security agencies the access they want.

That compromise will draw fire from both sides. Advocates of stricter frontier-model controls will argue that voluntary frameworks are worthless because labs can ignore them without penalty, and that cybersecurity risks are too severe to leave compliance optional. They’re not wrong—voluntary commitments are only as strong as reputational pressure and the threat of future regulation. If a lab skips the review process and ships a model that gets weaponized in a major attack, the political blowback could be severe enough to trigger the mandatory regime everyone’s trying to avoid.

Civil libertarians, meanwhile, will worry about expanded roles for security agencies like the NSA in model access and assessment. Giving intelligence agencies pre-release access to frontier models—even for 30 days of red-teaming—creates a potential pipeline for capability harvesting or surveillance tool development. The order doesn’t specify what happens to the models after testing, whether agencies retain copies, or what safeguards prevent mission creep. Those are real concerns, and the classified nature of the benchmarking process makes oversight harder.

Think of it like this: the order is a airbag, not a seatbelt. It’s designed to cushion the impact of a frontier model cyber incident, not prevent labs from shipping risky systems in the first place. Whether that’s enough depends on how reckless you think the labs are and how fast capabilities are advancing.

Three Things to Monitor as Agencies Write the Rules

The 60-day classified benchmarking process is where the real policy gets made. Watch for leaks or public statements from labs about whether they’ve been contacted, what thresholds agencies are proposing, and whether smaller players get carved out. If the criteria are too broad, every foundation model triggers review and the program becomes unworkable. Too narrow, and it misses models that pose real cyber risks.

The 30-day timeline for strengthening federal cyber defenses will produce a flurry of agency memos and procurement actions. That’s the low-hanging fruit—deploying AI tools for threat detection, automating patch management, using large language models to analyze security logs. It’s less controversial than the frontier model framework, but it’s also where agencies will experiment with AI in production environments and learn what actually works. Failures here will shape how much trust agencies place in the voluntary review program.

Industry response will telegraph whether this framework has teeth. If OpenAI, Anthropic, and Google publicly commit to participating in pre-release reviews, the order succeeds in creating a de facto standard. If labs stay silent or hedge, it signals they’re treating this as optional theater. The first major model release after the 60-day window—likely sometime in late summer 2026—will be the real test. Does the lab coordinate with NSA and Treasury, or does it ship and dare the administration to respond?

Source: Holland & Knight (summarizing White House EO)