The federal government has spent the last year systematically tearing down every piece of guardrail meant to keep artificial intelligence from spinning out of control. Then, a single software model terrified the Treasury Department so thoroughly that the administration scrambled to undo its own deregulatory agenda.
President Donald Trump signed an executive order establishing what the administration calls a voluntary framework to give federal agencies early access to advanced AI models. Under the new directive, tech companies will be asked to hand over their most powerful systems for national security and cybersecurity testing up to 30 days before they are released to the public.
Behind the scenes, this order is a direct response to a massive policy panic. The administration originally planned a broad 90-day review period, but the president abruptly canceled an Oval Office signing ceremony after tech executives and venture capitalists argued that a three-month delay would hand the technological advantage to China. The compromise is a watered-down 30-day window that is entirely voluntary, highlighting a deep contradiction. Washington is terrified of what these models can do, yet utterly powerless to compel Silicon Valley to show them the code.
The Mythos Panic
To understand why the administration reversed its hands-off approach to Silicon Valley, look at what happened in April. Anthropic announced a preview of its next-generation model, Claude Mythos. Unlike previous iterations that generate text or code on command, Mythos possesses an unprecedented ability to autonomously hunt for, identify, and exploit deep architectural vulnerabilities in real-world software systems.
It is a weapon-grade capability wrapped in a commercial wrapper.
Treasury Secretary Scott Bessent and officials from the Federal Reserve held urgent briefings with major Wall Street chief executives. The warning was stark: if Mythos or a competitor's equivalent were released without warning, the global financial infrastructure could be exposed to automated, AI-driven cyber warfare overnight. The defense systems protecting large banks, energy grids, and government networks are built on legacy code. A model that can map out thousands of structural flaws in seconds shifts the balance heavily in favor of attackers.
The government faced a crisis of its own making. Having dissolved previous oversight structures and neutralized the independent testing capacity of federal institutes, the administration had no formal mechanism to inspect the technology before it hit the market. The resulting executive order was not a planned strategic initiative. It was an emergency patch.
The Illusion of Voluntary Compliance
The core flaw of the new order rests on a single word: voluntary. The text explicitly states that nothing in the document authorizes the creation of a mandatory governmental licensing, pre-clearance, or permitting requirement.
[Private AI Labs: OpenAI, Anthropic, Google]
│
▼ (Voluntary 30-day window)
[Classified Benchmarking Process]
│
├─► Department of War / Defense
├─► Treasury Department
└─► Homeland Security
This structural limitation creates an environment where federal agencies are entirely dependent on the goodwill of commercial developers. If an AI lab decides that a 30-day delay hurts its competitive edge against its rivals, it can simply bypass the government review entirely.
National security officials argue that market pressure will force compliance. They believe no major tech giant will risk the reputational damage of releasing a model that later cripples a power grid if it is discovered they skipped the federal review. But history shows that commercial pressure almost always overrides caution. In a market where missing a launch window by two weeks can wipe billions off a company's valuation, a voluntary 30-day pause is an eternity.
Furthermore, the review process itself is deeply compromised by industry influence. Because federal agencies lack the specialized computing infrastructure and the elite engineering talent required to stress-test these behemoths, the order relies on an AI cybersecurity clearinghouse. This body will lean on the expertise of private industry and selected financial institutions to do the actual vetting. The government is essentially asking Silicon Valley to grade its own homework.
Internal Warfare in the Administration
The path to this executive order exposed a bitter ideological rift among the policymakers advising the White House.
On one side stood the safety contingent, championed by figures who pointed out that leaving national infrastructure exposed to autonomous exploitation is a failure of basic statecraft. Even Kevin Hassett, director of the National Economic Council, suggested behind closed doors that advanced frontier models should face an approval process modeled after the Food and Drug Administration. He argued that technology capable of systemic economic disruption should be proven safe before deployment.
On the other side was a powerful coalition of Silicon Valley donors and libertarian-leaning advisers, including venture capitalist David Sacks. They viewed any federal delay as a strategic surrender to foreign adversaries. Their argument won the first round, prompting the president to halt the original, stronger draft of the order because he feared it would blunt the American lead over international rivals.
The final text reflects a complete surrender to the anti-regulation camp. By cutting the review window from 90 days to 30, and ensuring it applies only to major generational leaps rather than incremental updates, the administration created an oversight mechanism that is largely performative. Thirty days is barely enough time for an agency to set up the server environments to run a model, let alone conduct a thorough, adversarial red-team analysis of its emergent capabilities.
The Looming Architecture of Licensing
While the administration insists this order is a victory for light-touch oversight, veteran policy analysts see a different trajectory. By defining what constitutes a covered frontier model through a new classified benchmarking process, the government is establishing the exact technical definitions needed for future regulation.
The order directs the Department of War, the Treasury, and Homeland Security to formalize these benchmarks. Once the government establishes a rigid framework for what makes an AI model inherently dangerous, the transition from a voluntary check-in to a mandatory licensing regime becomes a matter of political will. The next time a model causes a significant, unexpected market disruption or an automated infrastructure failure, the infrastructure for a full federal shutdown of unapproved software will already be sitting on the books.
For now, the tech sector is celebrating the order as a sign that they have successfully kept federal regulators at bay. But the reality is far more fragile. Washington has admitted it is terrified of what is being built in the labs of San Francisco, and this voluntary agreement is the last attempt to manage that fear through cooperation before the inevitable clampdown.
