On June 2, President Trump announced an executive order on the cyber risks posed by frontier AI models. Titled “Promoting Advanced Artificial Intelligence Innovation and Security,” the order addresses the administration’s concern that advanced AI capabilities “introduce new national security considerations that require coordinated action across executive departments.” This concern follows from recent evaluations of frontier AI models—specifically, Anthropic’s Mythos and OpenAI’s GPT 5.5—indicating heightened abilities to identify and exploit software vulnerabilities.
In response, the order calls for several actions to render the federal government and critical infrastructure entities more resilient in light of such threats. It also outlines a voluntary, 30-day pre-deployment testing process during which frontier AI models can be examined for novel capabilities—affording the government additional time to ready itself and key stakeholders for a subsequent general release.
This order reflects the administration’s evolving posture on how best to ensure the nation’s AI dominance. As noted in the order, this pre-deployment review and related resiliency efforts are indicative of the administration’s desire to “ensure that the best and most secure technology is deployed rapidly to confront any and all threats to our country.” Amid mounting public concerns over the reliability and safety of AI models and the trustworthiness of AI companies, creation of a transparent, evidence-based testing process of leading models may hasten both adoption and development of AI—a process that can be established without adopting more restrictive regulatory models of premarket approval that the administration was previously rumored to have been considering.
Yet, a review of the order’s key provisions demonstrates how it will likely fall short of mitigating the burgeoning techlash. The order raises red flags around a lack of transparency and specificity. Such opacity and ambiguity risk turning the order into a vehicle for excessive and arbitrary government interference.
Steps to Generally Bolster Cyber Readiness
The order calls on several departments to take various measures to shore up the nation’s defense systems and critical infrastructure as cyber risks introduced by AI become more pronounced. This includes the Department of Defense bolstering its information systems.
It also instructs the director of the Cybersecurity and Infrastructure Security Agency (CISA) to issue binding operational directives (BODs) related to protecting civilian information systems. Pursuant to the Federal Information Security Modernization Act of 2014, CISA may direct other agencies to take necessary actions to protect federal information and information systems from “known or reasonably suspected information threats, vulnerabilities, or risks.” Such directives have previously mandated that agencies patch critical vulnerabilities and update systems to include the latest safeguards. CISA is also tasked with facilitating access to cybersecurity tools to state and local authorities and operators of critical infrastructure in response to any new threats posed by frontier AI models.
To further streamline and expedite the process of proactively identifying and mitigating emerging cyber risks, the order sets a 30-day deadline for the secretary of the Treasury to form an “AI cybersecurity clearinghouse.” Participants in the clearinghouse may include AI industry stakeholders and operators of critical infrastructure. This new body will coordinate the process of scanning for software vulnerabilities and taking responsive action.
It’s likely that these steps will be followed by further cyber measures. The Office of Management and Budget has 60 days under the order to assess if any “available and relevant funding” can be allocated to this broader cyber resilience effort.
A Voluntary, Yet Opaque Pre-Deployment Testing Process
The order’s core provisions (as well as those drawing the most analysis) stand up a voluntary mechanism for evaluating frontier AI models. According to an X post by the Office of Science and Technology Policy (OSTP), it “creates a process for frontier labs to voluntarily share cutting-edge cyber models in order to secure critical infrastructure and strengthen the government’s own cyber defenses.” The OSTP clarified, “We are NOT conducting oversight of all new models, as that level of government overreach would have chilling effects on free speech and innovation.” A close reading of the order, however, does not foreclose potential abuse of this voluntary system.
Single officials exercise considerable discretion over key parts of this process. Authority over the creation of a “classified benchmarking process”—a technical means to identify models that surpass certain capabilities and therefore qualify as “covered frontier models”—is left to the director of the National Security Agency (NSA). While the NSA has permission to consult other departments, its role as the primary decisionmaker over such a consequential process stands out. With the NSA at the helm, this process will necessarily be more opaque and will likely be grounded in the values and priorities of the Intelligence Community (IC).
It’s also not immediately clear how much of this process needs to be classified. Anthropic, OpenAI, Google, and Meta publish thorough reports on the capabilities of their respective models. They omit some details as to how exactly they measured those models, but they commonly provide qualitative and quantitative evidence to help officials and the public understand what a model may be capable of when released. The more information that the NSA withholds, the more likely it is that the process will be scrutinized and questioned.
Questions around accountability and transparency also emerge from the order’s provisions related to the design of a voluntary framework for the developers of all such covered frontier models. The secretary of the Treasury and the secretary of Defense, through the NSA; the secretary of Homeland Security, through the CISA director; and the assistant to the president for Science and Technology Policy and the secretary of Commerce, through the director of the National Institute of Standards and Technology, are assigned this task.
The framework must allow AI developers to “engage” with the federal government to assess if their model qualifies as a covered frontier model. When they should engage and how long the federal government may take to perform that evaluation task is left open.
When and if a model does qualify, then the framework shall permit AI developers to share their models with the federal government “for a period of up to 30 days before they plan to release such models to other trusted partners.” David Sacks, former AI czar, reported on X that the order refers to 30 calendar days as opposed to business days—a critical difference given that “in the AI race, every day counts.” Even if Sacks is right on that front, the order may nevertheless expect labs to then go through additional steps prior to a general release. Following the 30-day period for government evaluation, the order implies that labs are then expected to go through another pre-deployment phase with trusted partners.
While the order makes clear that this process in no way permits creation of a mandatory process, it does not take an extensive imagination to see how this process could be used to place pressure on labs. Here’s a hypothetical scenario generated by the AI frontier model Claude:
Imagine a leading frontier lab—let’s call it NovaMind—quietly decides that the voluntary framework isn’t worth the operational cost. Its legal team calculates that 30 calendar days of prerelease government access, potentially followed by an extended trusted-partner phase, amounts to a soft embargo that competitors not yet at the frontier threshold can freely exploit. NovaMind says nothing publicly. It simply releases its next model, NovaMind‑7, on a Tuesday morning.
By Tuesday afternoon, the NSA director has briefed the national security advisor. By Tuesday evening, the White House chief of staff has alerted the communications team. By Wednesday morning, the OSTP account on X posts a terse statement: “NovaMind declined to participate in the voluntary cybersecurity review process established by the president. Critical infrastructure operators and federal agencies were given no advance notice that a model with significant cyber-offensive capabilities had been released into general circulation. We are assessing the implications.”
The post gets 40,000 reposts in six hours. The phrase “NovaMind refused to cooperate” trends nationally. Three senators—two from the Judiciary Committee and one from Armed Services—issue statements questioning whether NovaMind’s government contracts should be reviewed. A Department of Defense spokesperson notes, without elaboration, that all vendors are expected to “support the national security mission.” No formal action is taken. None is needed. The message has been delivered through the loudest possible megaphone at zero legal cost to the government.
NovaMind‑8 enters the voluntary review process 14 weeks later.
In conclusion, the order identifies a genuine concern around national security and the resilience of our critical infrastructure. However, it responds with a framework that is, at best, half-built. Benchmark classification authority rests predominantly with the IC behind a classified curtain, with no procedural floor for response times and no public methodology to scrutinize. The NovaMind scenario is not paranoid speculation. It is a straightforward application of how informal government pressure may work in practice, and the order’s disclaimer does nothing to foreclose it.
Congress has the tools to devise a more durable, specific, and transparent approach: Public benchmarking standards, defined response timelines, and an explicit prohibition on contract retaliation against nonparticipating labs would convert a presidential aspiration into a framework more aligned with the rule of law.