Trump signs AI executive order asking companies to let the NSA test frontier models before release

President Trump signed an executive order on June 2, 2026, titled "Promoting Advanced Artificial Intelligence Innovation and Security," establishing a voluntary framework that asks AI companies to give the federal government early access to frontier models for cybersecurity testing before those models are released to the public.

The order represents a significant shift for an administration that initially took a hands-off approach to AI regulation. It directly revives elements of an earlier proposal that the White House shelved after industry pushback — but with a key change: where the original draft would have required a 90-day mandatory review period, the final order asks for voluntary participation only.

What the order actually does

The core mechanism is a request — not a mandate — for AI companies to allow federal agencies to evaluate their "covered frontier models" up to 30 days before public release. These are defined as the most powerful AI models, specifically those capable of identifying and exploiting software vulnerabilities at scale. Participation is explicitly voluntary, and the order expressly prohibits the creation of any mandatory governmental licensing or permitting requirements for AI development or distribution.

Supporting the voluntary testing framework, the order establishes an AI cybersecurity clearinghouse coordinated by the Treasury Department, the NSA, and CISA. The clearinghouse is tasked with scanning for software vulnerabilities at scale, validating discovered vulnerabilities, and prioritizing patches — operating in coordination with the AI industry and critical infrastructure operators.

The NSA is given a specific additional role: developing a classified benchmarking process to assess the advanced cyber capabilities of AI models. That benchmark will also define the threshold at which a model is designated a "covered frontier model" subject to the voluntary review request.

Industry shape and criticism

The final order is notably softer than its predecessor. An earlier draft, reportedly circulated in May, included a 90-day mandatory review window that drew significant opposition from AI companies concerned that pre-release government access would slow development timelines and expose proprietary model architecture. The June 2 order dropped the mandatory language entirely.

The classified nature of the NSA's frontier model threshold drew criticism from Dean Ball, a former Trump administration AI policy adviser, who argued the public and research community have a right to understand how the government defines models capable enough to warrant security review. The White House described the classified threshold as necessary to prevent adversaries from calibrating their own models to avoid triggering government scrutiny.

Congressman Josh Gottheimer, who has pushed for stronger AI oversight legislation, said the voluntary framework lacks the "teeth" to ensure AI companies actually participate. The order's effectiveness, as several analysts noted, depends entirely on whether leading AI labs — Anthropic, OpenAI, Google DeepMind, Meta, and others — choose to submit models before release.

Context: why the administration moved at all

The order's stated impetus is the demonstrated capability of frontier AI models to autonomously discover software vulnerabilities. The White House fact sheet referenced Anthropic's Project Glasswing, in which Claude Mythos Preview reportedly identified over 10,000 high- or critical-severity vulnerabilities in systemically important software within a single month. That kind of capability, in the hands of adversaries, represents a new category of risk to critical infrastructure that the administration concluded it could not ignore.

The executive order directs federal agencies to expand their use of AI-enabled defensive cybersecurity tools and to make those tools available to state and local governments and critical infrastructure operators — positioning AI as both the threat and the primary defense against that threat.

Whether the voluntary framework produces meaningful pre-release reviews will depend on whether the AI industry sees participation as worth the access costs. The administration left that question deliberately open.

Sources: White House; The Next Web; Council on Foreign Relations