The landscape of artificial intelligence regulation in the United States is on the cusp of a significant shift. According to a report from the Financial Times, the Trump administration is within days of finalizing a voluntary standards agreement with several of the nation's largest AI companies. This deal, expected to be announced “as early as next week,” would establish a set of benchmarks for frontier AI models, particularly concerning their cybersecurity capabilities. The report, which cites anonymous sources familiar with the negotiations, suggests that the agreement could provide much-needed clarity in a regulatory environment that has been anything but clear.
For months, AI companies have operated in a grey area. The rules governing what these models can and cannot do have been ambiguous, leaving firms vulnerable to sudden government actions. The most notable recent example is the export control directive delivered to Anthropic on June 12, which effectively shut down its latest publicly released model for the remainder of June. Anthropic, a San Francisco-based company founded by former OpenAI employees, had been pushing the boundaries of AI safety research. Its model, Claude, had been seen as one of the most capable and safety-conscious offerings on the market. Yet, the government's intervention sent shockwaves through the industry. OpenAI, clearly rattled by the move, subsequently withheld the release of its own latest models, likely as a precautionary measure to avoid a similar fate. These events underscore the urgent need for a predictable regulatory framework—one that companies can rely on without fear of sudden crackdowns.
The forthcoming deal is expected to be a voluntary agreement, meaning companies will not be legally compelled to adhere to the standards. However, the implicit threat of government action, as demonstrated with Anthropic, suggests that non-compliance could carry significant risks. The agreement will be overseen by the Center for AI Standards and Innovation (CAISI), housed within the Commerce Department, and the National Security Agency (NSA), which falls under the Pentagon. These two bodies will be central to defining and enforcing the benchmarks. One of the anonymous sources cited by the Financial Times emphasized that the NSA's involvement is crucial because the standards are primarily focused on cybersecurity—an area where the intelligence community has deep expertise.
The shift in the administration's stance is remarkable. At the beginning of President Donald Trump's second term, Vice President J.D. Vance signaled a laissez-faire approach to AI regulation. The prevailing sentiment was one of minimal government interference, allowing innovation to flourish without bureaucratic red tape. But that changed dramatically with a series of executive orders and direct actions against companies like Anthropic. The latest initiative appears to be the formalization of some of those earlier measures. The government's stated goal, according to a previous executive order, is 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.” This language is significant for several reasons.
First, the mention of a classified benchmarking process means that the public will not have access to the specific criteria used to evaluate AI models. This raises immediate concerns about transparency and accountability. Civil liberties groups and some lawmakers have already expressed unease about the lack of oversight in such a critical area. Without public knowledge of the standards, there is no way to independently verify whether the government is overreaching or whether companies are being held to reasonable expectations. Second, the designation of a “covered frontier model” suggests a tiered regulatory approach, where only the most powerful and advanced AI systems will be subject to these standards. This could leave smaller players and open-source models outside the scope of scrutiny, potentially creating a loophole for less responsible actors.
The companies reportedly involved in the discussions include Anthropic, OpenAI, Amazon, Microsoft, and Google. Notably absent from the list is Meta, the parent company of Facebook. According to other anonymous sources who spoke to the press about a week ago, Meta has been a holdout in the negotiations. The Trump administration has reportedly been working “overtime” to get Meta's buy-in, but the company's reluctance may stem from its heavy investment in open-source AI models. Meta's Llama models are among the most widely used open-source language models, and the company has advocated for a more permissive regulatory environment that encourages widespread distribution. Subjecting these models to classified cybersecurity benchmarks could undermine Meta's open-source strategy, potentially requiring the company to lock down its models in ways it has not previously contemplated.
The historical context of AI regulation in the United States is fraught with tension between innovation and safety. During the first Trump administration, the focus was overwhelmingly on maintaining American leadership in AI technology. The American AI Initiative, launched by executive order in 2019, emphasized removing barriers to innovation and ensuring that regulations were not overly burdensome. The Biden administration, by contrast, took a more cautious approach, issuing an executive order on “Safe, Secure, and Trustworthy Artificial Intelligence” in 2023, which imposed new reporting requirements and pushed for safety testing. The Trump administration's second term, however, initially signaled a return to a deregulatory posture—until the actions against Anthropic showed that the government was willing to wield its power.
The current deal, if finalized, could represent a middle ground. A voluntary agreement allows the administration to claim it is taking action on AI safety without imposing the kind of mandatory regulations that might alienate the tech industry. For the companies involved, it provides a measure of predictability and a seat at the table in shaping the standards. Yet, the voluntary nature also raises questions about enforcement. What happens if a company fails to meet the benchmarks? The report does not specify consequences, but the precedent set by the Anthropic case suggests that the government is prepared to use export controls or other tools to compel compliance.
From a technical perspective, the cybersecurity benchmarks are likely to focus on preventing AI models from being used to automate cyberattacks, develop sophisticated malware, or exploit vulnerabilities in critical infrastructure. The NSA's involvement underscores the national security implications. Advanced AI models have demonstrated the ability to write convincing phishing emails, identify security flaws in code, and even generate functional exploits. As these capabilities improve, the potential for misuse grows. The classified benchmarks would presumably set thresholds for when a model's cyber capabilities become too dangerous to be released publicly without safeguards. This could include requirements for red-teaming, input filtering, and access controls.
Critics argue that classified benchmarks are contrary to the spirit of open scientific inquiry that has driven AI progress. Many of the most important advances in AI have come from academic institutions and open collaborations, not just from corporate labs. Secrecy around standards could stifle academic research and create an uneven playing field where only a few large companies have access to the full regulatory landscape. Moreover, the international dimension cannot be ignored. China and the European Union are also developing their own AI governance frameworks. A classified U.S. approach could complicate international cooperation on AI safety, as other nations may be reluctant to share sensitive information if the U.S. standards are not transparent.
The timeline for the announcement is still uncertain, but the Financial Times report indicates that it could come as early as next week. The exact list of participating companies may be clarified at that time, as well as the specific areas covered by the standards. In addition to cybersecurity, there is speculation that the agreement may include provisions for watermarking AI-generated content and preventing the use of models for disinformation. However, the primary focus appears to be on cyber capabilities, reflecting the national security priorities of the NSA and the Trump administration.
For the AI industry, this deal represents both an opportunity and a challenge. On one hand, it offers a clear path forward for companies that have been navigating a fog of uncertainty. On the other hand, the voluntary nature and classified details mean that the public will have limited visibility into what is being agreed upon. Whether this is ultimately a good thing or a bad thing depends on one's perspective. Supporters will argue that it strikes the right balance between safety and innovation, keeping the government involved without imposing heavy-handed regulations. Detractors will point to the lack of transparency and the potential for abuse by powerful interests behind closed doors.
As the world watches, the next few days will reveal whether this deal becomes a template for AI governance in the United States. The stakes are high, and the outcome will shape the development of artificial intelligence for years to come. Whatever happens, one thing is clear: the era of completely unregulated frontier AI development is coming to an end, but the nature of the new regulatory framework remains very much in flux.
Source:Gizmodo News