The Trump administration released a long-anticipated executive order this month aimed at exercising oversight of cybersecurity risks posed by the most advanced artificial intelligence models. The order attempts to thread the needle between the administration’s anti-regulatory posture and the recognition that it cannot remain on the sidelines in the face of AI’s serious risks to national security.
Last week’s chaotic federal ban on foreigners from using Anthropic’s new Claude Fable 5, however, may have provided a preview of how the administration’s regulation of AI will play out, and many questions remain.
Fable 5 is the publicly accessible version of Claude Mythos 5, a large language model the company released in April 2026 that is capable of identifying security flaws in software systems. The model can help governments, banks, utilities, hospitals, and other providers of essential services detect and patch the holes. But bad actors could also use this capability to exploit these weaknesses to freeze payment networks, shut down hospitals, and take telecommunications and internet services offline.
The release of Mythos was part of the motivation behind President Trump’s order, which establishes a voluntary process for AI developers to submit their models to the government for classified testing for a period of up to 30 days, after which they can release the model to other trusted partners. The order also creates an “AI security clearinghouse” that fosters collaboration between AI developers and operators of critical infrastructure to identify and patch software vulnerabilities.
The administration’s newfound openness to AI safeguards will not by itself undo the evisceration and politicization of regulatory agencies that are now expected to shoulder the responsibility of testing, mitigating, and overseeing a rapidly evolving threat landscape. Testing and threat sharing alone will not head off a cybersecurity catastrophe. Achieving that will also depend on how the administration revives and mobilizes regulatory expertise across the government. And it will have to collaborate with industry players to establish guardrails on emergent AI capabilities while maintaining leverage to slow or halt their release as needed to shore up the nation’s defenses and critical infrastructure.
In other words, the order’s success will turn on how effectively the administration engages in the high wire act of governing, which the administration can’t do alone. Congress must also step in to establish testing standards and a clear structure of leadership and accountability. The Brennan Center will be tracking at least three questions related to the order.
Who in government will be responsible for the testing process?
Three agencies are jointly responsible for setting up the classified testing process, a multi-stage assessment known as benchmarking: the Cybersecurity and Infrastructure Security Agency (CISA) within the Department of Homeland Security, the Treasury Department, and the National Security Agency, which has a dedicated cybersecurity directorate.
But CISA, the agency focused on addressing threats to the nation’s physical and cyber infrastructure, is in tatters. The Trump administration has reduced its staff by a third and forced out nearly all of its senior career officials. The agency has been without a Senate-confirmed leader since Trump took office, and the White House’s latest budget proposal would slash its budget further by $700 million. The NSA has largely been spared from steep cuts but faced a politically motivated purge of its leadership last year that could hamper its capacity to coordinate implementation of the order’s testing and oversight frameworks.
These cuts have wiped out decades of in-house expertise on cybersecurity threats and sapped morale. The order does not say how either agency will be staffed and resourced to test and safeguard the cyber capabilities of advanced AI models, which is a fraught and complex undertaking even under the best of circumstances. Instead, the order appears to look to the Treasury Department to fill the leadership vacuum. While familiar with the security threats to banking infrastructure, the Treasury Department lacks the expertise needed to helm a whole-of-government approach to risks from AI.
The Commerce Department has also asserted expansive authority to regulate access to AI models, even though the order limits it to a consultative role in developing the classified testing process. Last Friday, the department’s Bureau of Industry and Security directed Anthropic to suspend access to Fable 5 and Mythos 5 by any foreign national under the bureau’s power to establish export controls on sensitive technology, apparently because another company tipped it off to the possibility that foreign adversaries could exploit a loophole in the models’ restrictions on identifying software vulnerabilities. But this decision came before the order’s classified testing process could even have been rolled out, injecting even more unpredictability into how the administration will act on information gleaned from and outside its own testing.
Who will have the final say over whether a model is safe for release?
Testing the cyber capabilities of AI models will rarely be straightforward. Agency testers might, for example, reach different conclusions from AI developers about the scope and severity of the vulnerabilities they have identified. The two sides might also disagree about the best ways to remedy these vulnerabilities. And even if there were consensus, the needed fixes will take time to roll out. That’s a lot to resolve in the span of 30 days, the maximum window afforded to the government to test models before they can be released broadly. A previous version of the order had a 90-day review period, but Trump nixed the draft after raising concerns that it would stifle innovation.
It’s anybody’s guess what might happen if the 30-day period lapses and the government has not completed its benchmarking evaluations or if it remains at odds with AI developers about how to fix issues discovered during the review process. The voluntary nature of the testing arrangement means AI developers can, it would seem, override the government’s concerns and release their models. Developers are well aware, however, that they could face legal and financial backlash if they elect to move forward without the government’s blessing. And the administration’s sudden decision to lock down access to Anthropic’s models based on vaguely specified national security concerns indicates that it might even preempt its own testing window.
Anthropic’s bruising fight with the Department of Defense has also put industry on notice about the lengths this administration will go to strongarm those that refuse to do its bidding. In February, the Department of Defense designated Anthropic as a “supply chain risk” after the company refused to drop contract restrictions that would limit Claude’s use in domestic mass surveillance and in fully autonomous lethal weapons — a designation that a federal judge concluded was an unconstitutional abuse of executive power. The Commerce Department’s restriction of Fable 5 has raised questions about whether Anthropic’s feud with the Pentagon might have played a role, though allies of the administration have denied this.
