The Trump administration is reportedly on the verge of announcing a voluntary standards deal with several of the largest frontier AI companies in the United States. According to the Financial Times, citing anonymous sources familiar with the negotiations, the agreement could be unveiled as early as next week. The deal is expected to establish a set of shared standards for frontier AI models, particularly focusing on cybersecurity capabilities. This marks a significant shift from the administration’s earlier laissez-faire approach to AI regulation and represents one of the most concrete steps yet toward formalizing oversight of the rapidly advancing technology.
The reported deal involves key government agencies: the Center for AI Standards and Innovation (CAISI), housed under the Commerce Department, and the National Security Agency (NSA), which operates under the Pentagon. These bodies would be central to implementing and monitoring the standards once formalized. The standards are described as voluntary, meaning companies would not be legally bound to comply, but the administration has already demonstrated its willingness to take coercive measures. In June, the U.S. government issued an export control directive to Anthropic that effectively shut down the company’s latest publicly released model for the remainder of the month. OpenAI, evidently wary of similar action, has withheld the release of its newest models as a precautionary measure. This context underscores that while the deal may be framed as voluntary, the threat of unilateral government action looms large.
The White House’s executive order on artificial intelligence, issued earlier this year, laid the groundwork for this development. The order calls for the government to “develop and maintain a classified benchmarking process to assess the advanced cyber capabilities of AI models and determine the threshold at which an AI model should be designated a ‘covered frontier model’ for the purposes of this order.” It also requires sharing such assessments with AI developers and researchers as appropriate. The classified nature of this benchmarking process means that the public may never know exactly what standards these companies are being held to. However, the shared practices across multiple firms—such as common safeguards and reporting requirements—could offer some indirect transparency.
The companies reportedly involved in the negotiations include Anthropic, OpenAI, Amazon, Microsoft, and Google. Notably absent from the list is Meta, which according to earlier leaks from anonymous sources has been a holdout in the talks. The Trump administration has reportedly been working overtime to secure Meta’s buy-in, but as of the time of reporting, no deal with the social media giant has been announced. This raises interesting questions about the competitive dynamics at play. Meta’s open-source approach to AI, embodied by its Llama models, may clash with the more controlled, security-focused framework that the administration seems to prefer. The absence of Meta could also create a two-tier system where some major players operate under the voluntary standards while others do not, potentially leading to fragmentation or regulatory arbitrage.
The shift in the Trump administration’s AI policy has been dramatic. At the start of the second Trump term, Vice President J.D. Vance signaled a hands-off, pro-innovation stance, suggesting that the government would avoid heavy regulation to allow American AI companies to compete globally. However, a series of actions—including the executive order, the crackdown on Anthropic, and now this standard-setting initiative—have illustrated a pivot toward active governance. This change reflects growing concerns about the national security implications of advanced AI, particularly in areas such as cybersecurity, disinformation, and critical infrastructure protection. The involvement of the NSA underscores the security lens through which the administration now views frontier AI models.
The classified benchmarking process is particularly controversial. While proponents argue that secrecy is necessary to prevent adversaries from gaming the system, critics warn that it undermines accountability and public trust. Without knowing the specific metrics by which models are judged, it becomes difficult for independent researchers, civil society organizations, and the broader public to assess whether the standards are rigorous enough—or whether they are too lenient. Moreover, classified benchmarks could be used to shield government decisions from judicial review, creating a black box around AI regulation.
Industry reactions to the potential deal are mixed. Some executives welcome clarity after months of uncertainty about what rules apply. The voluntary nature of the agreement offers flexibility, allowing companies to adapt without being locked into rigid legal mandates. Others, however, express concern that the classified requirements could stifle innovation and create barriers for smaller players who lack the resources to comply with opaque standards. There is also the risk that the deal could be used to entrench the dominance of a few large firms, as compliance costs and security demands may favor established players with deep pockets.
From a historical perspective, this development fits a pattern of the U.S. government gradually building infrastructure for AI governance. Previous efforts, such as the National AI Initiative Act of 2020 and the establishment of the National Institute of Standards and Technology’s AI Risk Management Framework, have laid institutional groundwork. The creation of CAISI under the Commerce Department is a direct outgrowth of this. But the current deal goes further by tying voluntary standards to a classified process and involving the NSA—a move that blurs the line between commercial regulation and national security. This convergence has precedents in other technology sectors, such as telecommunications and encryption, where the government has historically pressed for backdoors and surveillance capabilities.
The timeline for the announcement is unclear, but sources suggest it could come within days. If finalized, the deal would represent the first major multilateral agreement on AI standards between the U.S. government and industry. Its success or failure could set a template for future regulation, not just in the U.S. but worldwide. International partners, particularly in Europe and Asia, are likely watching closely. The European Union’s AI Act, which adopts a risk-based approach with mandatory requirements for high-risk systems, contrasts with the U.S.’s largely voluntary framework. How the U.S. deal performs could influence whether other countries lean toward mandatory legislation or follow a softer path.
One key area of focus in the standards will be cybersecurity. Frontier AI models have capabilities that could be exploited by malicious actors to automate hacking, generate phishing attacks at scale, or discover vulnerabilities faster than humans. The NSA’s involvement suggests that the standards will emphasize model defenses against such threats, possibly including red-teaming requirements, access controls, and reporting obligations for certain capabilities. The classified benchmarks likely define thresholds beyond which models are deemed too dangerous to release without safeguards. This mirrors internal practices at companies like OpenAI and Anthropic, which already conduct pre-deployment safety evaluations. Government alignment could standardize these practices across the industry.
While the deal is still under negotiation, its outlines signal a new era in AI governance. The Trump administration, which once championed deregulation, is now embracing a form of oversight that marries security concerns with industry partnership. The voluntary label may soften the political blow for free-market advocates, but the substance of the agreement is deeply regulatory. For good or ill, the era of unrestricted AI development in the United States appears to be drawing to a close. As the details emerge, the public and policymakers alike will have to grapple with the implications of a system where the most powerful AI models are subject to standards that neither the press nor the people are allowed to fully scrutinize.
Source:Gizmodo News