The Trump administration is reportedly expanding its secretive AI safety framework to cover advanced open-weight models, a move that could reshape the balance between innovation and security in American artificial intelligence. According to a new report, the framework currently applies only to closed models from top-tier providers such as OpenAI. But officials are now preparing to include open models that are deemed advanced enough to warrant government inspection.

What We Know About the Secretive Framework

The framework itself remains largely classified. Reports indicated earlier this month that the administration had been assembling an AI safety framework to assess the cyber capabilities of models, but that the details would be kept secret from everyone except model providers. That secrecy has drawn criticism from transparency advocates, who argue that any government review of AI systems should be subject to public oversight. The reported expansion to open models adds a new layer of complexity, because open-weight models can be downloaded, modified, and redistributed by anyone.

According to the reporting, the initial version was built with closed models in mind. Companies like OpenAI, Google, and Anthropic would submit their most capable systems for review, and the administration would provide some form of official approval or clearance. But the administration is concerned that granting official approval only to closed models would harm the reputation of open models. Officials worry that a government seal of approval for closed systems, with no equivalent for open systems, would "paradoxically disincentivize" American companies from releasing open models at all.

The framework reportedly focuses on cyber capabilities, meaning the ability of an AI system to find vulnerabilities, write malicious code, or conduct offensive operations. These are the kinds of capabilities that national security officials worry about most. A closed model can be monitored while it stays in the hands of a trusted company, but an open model can be downloaded by anyone, including hostile state actors. That makes the prospect of government review far more complicated.

What Does "Open Weight" Mean?

An open-weight model is an AI system whose trained parameters, or weights, are made publicly available. Unlike a closed model, which can only be accessed through an API or proprietary platform, an open-weight model can be downloaded and run locally. Developers can fine-tune the model on their own data, integrate it into their own applications, and study its behavior. This openness is what makes these models so attractive to startups, academics, and governments around the world.

Openness is not the same as fully open source, however. Some open-weight models come with restrictive licenses that limit commercial use or impose other conditions. But the defining feature is that the weights are public, which means the model can be inspected and modified in ways that closed models cannot.

Why Open Models Are Suddenly in the Spotlight

Open models have gained renewed attention in the American AI world after a series of releases from Chinese laboratories shook the industry. Last month, Moonshot released Kimi K3, an open-source model that was significantly cheaper than comparable American systems while matching or exceeding the performance of some of the latest offerings from Anthropic, OpenAI, and Google. The release set off a wave of hand-wringing among U.S. AI executives and policymakers, many of whom had assumed that American closed models would remain ahead of everything else for years to come.

Kimi K3 demonstrated that Chinese open models are no longer just cheaper imitations. They are competitive on quality, and because they are open weight, developers around the world can adapt them for their own use. That has led to fears that Chinese open models could become the default foundation for AI applications outside the United States, especially in countries that do not want to rely on American cloud services. It has also raised national security questions about whether open models from rival nations could be used for cyberattacks, disinformation, or other malicious purposes.

Experts initially theorized that the Trump administration would respond to the Chinese threat with blanket bans or export controls. That did seem to be the main strategy for some time, according to people familiar with the administration's thinking. But the strategy shifted after AI industry leaders decided to make open models a cause worth fighting for.

Industry Pushback and the Open Letter

Numerous AI companies, including Meta, Microsoft, and Palantir, signed an open letter titled "Open Weights and American AI Leadership." The letter urged the Trump administration to avoid restricting open models, arguing that such restrictions would only cause the United States to fall further behind China. The signatories claimed that closed models are "not inherently safe" and that open development is essential for innovation, transparency, and national competitiveness.

The letter represented a rare moment of unity among companies that are often fierce competitors. Meta has long been one of the biggest advocates of open-weight AI, releasing models like Llama to a global audience. Microsoft has invested heavily in OpenAI but has also supported open-source initiatives. Palantir, known primarily for government and defense contracts, added an unusual voice in favor of openness, suggesting that the national security community itself sees value in open models.

Supporters of the letter argued that restricting open models would not make the United States safer. Instead, they said, it would cede leadership to Chinese researchers, who are already publishing open models without Western restrictions. They also pointed out that open models allow security researchers to inspect the underlying code and identify vulnerabilities, whereas closed models can hide flaws behind corporate secrecy.

Nvidia's High-Profile Intervention

Nvidia CEO Jensen Huang, who has long championed the idea that open models will be key to winning the global AI race, decided to make his first-ever X post on the subject. "Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty," Huang wrote. "The world needs both frontier closed models and frontier open models."

Huang had previously made the case to the Trump administration in private meetings, but the public post signaled a new willingness to engage in the policy debate directly. Nvidia's passion for the issue extended to the formation of the Open Secure AI Alliance, a coalition meant to aid the proliferation of open-weight AI models. The alliance brings together companies, researchers, and policymakers who share the goal of ensuring that open models remain a vital part of the global AI ecosystem.

Nvidia has a direct financial stake in the outcome, as open models can be run on Nvidia GPUs just as easily as closed models. But Huang has also made a broader argument: that the United States cannot win the AI race if it ties one hand behind its back. He has repeatedly said that open models are essential for the long-term health of the industry, and that attempts to restrict them would only benefit competitors in China and elsewhere.

The Fracture Among AI Leaders

Not everyone in the AI industry signed the letter. OpenAI initially declined to join, but CEO Sam Altman later added his co-signature. Anthropic refrained from backing the letter, which caused the pro-open model camp to accuse CEO Dario Amodei of trying to push the administration toward a stricter stance in favor of closed models. These accusations were fueled by Anthropic's business interests, as the company has become one of the leading developers of closed AI systems.

Anthropic had a very public falling out with the Trump administration not too long ago, and its relationship with Washington remains complicated. Critics claimed that any policy discouraging the proliferation of open models would only help Anthropic's bottom line, since fewer open competitors would mean more demand for closed models. Amodei has rejected the accusation, insisting that his position is based on safety concerns rather than commercial advantage.

The split among AI leaders highlights a deeper philosophical divide. Proponents of open models argue that transparency is essential for safety, because independent researchers cannot audit closed systems. Proponents of closed models argue that openness is precisely what makes a model dangerous, because malicious actors can remove safety filters and fine-tune the model for harmful purposes. Both sides claim to hold the high ground, and both sides have credible arguments.

Government regulators are caught in the middle. On one hand, they want to encourage American innovation and leadership. On the other hand, they want to prevent the most capable AI systems from falling into the wrong hands. The reported expansion of the secretive framework to include open models appears to be an attempt to split the difference: allow open models to exist, but subject the most advanced ones to government review before release.

Key Facts at a Glance

• The Trump administration is reportedly expanding its AI safety framework to include advanced open-weight models.
• The framework currently applies only to closed models from providers like OpenAI.
• It is designed to assess cyber capabilities but remains classified except to model providers.
• Officials worry that excluding open models would "paradoxically disincentivize" American companies from releasing them.
• Chinese open models such as Moonshot's Kimi K3 have matched or exceeded some American closed models.
• Meta, Microsoft, Palantir, and others signed an open letter urging the administration to avoid restricting open models.
• Nvidia CEO Jensen Huang made his first-ever X post to argue in favor of open models.
• OpenAI's Sam Altman later co-signed the letter; Anthropic has not.

What remains unclear is how the administration will define "advanced enough" and how the review process will work in practice. Will open models need to be submitted before their weights are published? Who will do the assessment? What happens if a model fails? None of these questions have been answered publicly, and the classified nature of the framework makes it unlikely that they will be anytime soon. For now, the only thing that seems certain is that the debate over open versus closed AI is no longer just a technical

Source:Gizmodo News