Key Takeaways
- President Trump delayed signing an executive order that would have created a 90‑day voluntary testing and vetting regime for frontier AI models, citing concerns it could hurt U.S. competitiveness against China.
- The draft order would have let the NSA conduct classified evaluations, the Treasury set up information‑sharing agreements, and multiple cybersecurity‑focused agencies define the scope of models covered.
- It aimed to formalize and deepen existing cooperative testing practices between the U.S. government and AI firms, moving from voluntary “hand‑holding” to more robust, government‑led assessments.
- Although the Trump administration previously dismissed AI safety policies as inhibitive, officials acknowledge the need for some guardrails to address emergent national‑security risks.
- Experts note that AI’s dual‑use nature is evident in military applications (drone warfare, targeting, command‑and‑control) and cybersecurity threats, where models can both defend and enable sophisticated attacks.
- Congressional views are shifting; some lawmakers who once opposed autonomous AI kill decisions now worry adversaries could gain an edge if the U.S. imposes stricter human‑in‑the‑loop requirements.
President Trump’s Decision to Delay the AI Executive Order
President Donald Trump announced he would postpone the release of an executive order establishing a 90‑day testing and vetting regime for frontier AI models, just hours before the White House planned to publicize the signing. Speaking to reporters in the Oval Office, Trump said he delayed the order because he “didn’t like certain aspects of it” and worried it could undermine U.S. AI industry competitiveness, particularly vis‑à‑vis China. The move reflects a continuation of his administration’s skeptical stance toward AI safety measures that it previously argued would stifle innovation.
Core Provisions of the Draft Order
According to multiple sources, a draft of the order circulating in the previous 24 hours would have created a voluntary testing framework between the federal government and leading AI companies. Under this regime, the government would study new frontier models for up to 90 days before their public release. In addition to direct government evaluation, the draft would have facilitated access to these models for cybersecurity testers operating in critical‑infrastructure sectors such as finance and healthcare, aiming to uncover vulnerabilities before deployment.
Agency Roles and Responsibilities
The draft empowered the National Security Agency (NSA) to conduct classified evaluations of frontier AI models, while the Department of the Treasury would have established a new information‑sharing agreement linking AI firms with cybersecurity defenders in critical infrastructure. Other agencies—including the Office of the National Cyber Director, the Cybersecurity and Infrastructure Security Agency (CISA), and the National Institute for Standards and Technology (NIST)—were slated to help define which models fall under the vetting regime and to develop standards for testing and assessment.
Building on Existing Cooperation
In essence, the order sought to formalize an already cooperative relationship between AI developers and governments like the United States and the United Kingdom. Tech‑focused agencies and regulators have previously been granted access to earlier models ahead of public release for testing and evaluation. The draft aimed to transition from informal, ad‑hoc arrangements—often reliant on AI companies’ self‑explanations—to a more structured, government‑led process capable of deeper scrutiny.
Insights from a Former Federal Official
A former federal official who had seen the latest draft told CyberScoop that, based on conversations with the administration, the order was intended to enable “more robust testing from government agencies” compared with the limited, containerized reviews previously conducted for earlier models. The official, speaking anonymously, noted that past evaluations involved considerable “hand holding” from AI companies, wherein firms explained what they expected their models to do. The new approach would have allowed the government to receive the models outright and conduct independent assessments, reflecting a maturation in federal understanding of AI over the past five years.
Trump Administration’s Shift in Tone
The delay represents a stark pivot for the Trump administration, which entered office openly dismissive of AI safety policies, arguing they would inhibit U.S. industry. Trump’s remarks echo that earlier skepticism, framing the order as potentially harmful to competitiveness. Yet the former official observed that the administration’s early rhetoric has “painted them into a corner”: having rejected the idea of institutional guardrails, they now lack a better term for ensuring that emergent frontier models do not disrupt security, resorting to the language of testing and guardrails as a compromise.
National‑Security Concerns Driving the Debate
While debate continues over the optimal way to regulate AI‑related harms, there is broad consensus that genuine national‑security risks surround the technology. Experts warn that frontier AI models can be weaponized for military applications, cyber operations, and surveillance, imbuing engineering decisions with life‑or‑death consequences. The administration’s hesitation highlights the tension between fostering innovation and mitigating threats that could undermine national security.
Evolution of AI Red‑Team Practices
Ram Shankar Siva Kumar, founder of Microsoft’s AI red team, illustrated how the field has evolved. In 2019 his team comprised only himself and a few security and machine‑learning specialists. Today it includes a larger cadre of technologists supported by experts in psychology, linguistics, bioweaponry, and other disciplines. Siva Kumar noted that the emergence of “frontier harms” has necessitated this multidisciplinary expansion, as teams now anticipate a wider array of malicious uses beyond traditional software bugs.
Military Integration of AI Technologies
The United States, alongside Israel, Russia, Ukraine, and other nations, has already deployed AI in targeted military operations or integrated it into broader command‑and‑control structures. AI is being used to enhance drone warfare, power global hacking campaigns, and refine surveillance and targeting of both military personnel and civilians. These applications mean that decisions made by frontier AI companies can directly affect battlefield outcomes, raising ethical and strategic questions about autonomous lethal systems.
Congressional Reconsideration of Autonomous Kill Decisions
Some congressional members who previously opposed allowing AI to make autonomous kill decisions on the battlefield are now reassessing their stance. Rep. Don Beyer (D‑Va.), who co‑chaired the Congressional AI Caucus and served on a bipartisan AI task force in 2024, acknowledged the need to guard against dehumanizing such decisions but warned that adversarial nations might not impose similar constraints. Beyer cautioned that if the United States insists on a human‑in‑the‑loop while rivals do not, the non‑human approach could consistently prevail, putting U.S. forces at a disadvantage.
Cybersecurity Implications of Advanced AI Models
Experts have grown increasingly concerned about AI’s impact on cybersecurity. Current models excel at identifying software bugs and vulnerabilities, while newer systems such as Anthropic’s Mythos and OpenAI’s Daybreak can chain multiple exploits to launch more sophisticated attacks. Although state‑sponsored hackers experiment with AI to gain targeted efficiencies, private‑sector and law‑enforcement cybersecurity professionals observe that the technology has largely benefited cybercriminals and scammers, enabling them to automate and scale illicit activities with unprecedented precision. This dual‑use nature underscores why policymakers continue to grapple with how to harness AI’s advantages while curbing its potential for harm.