Trump's AI Safety Framework Skips Open-Source Models Entirely

Key points

• The Trump administration's AI cybersecurity testing framework, finalized in mid-2025, excludes all open-source AI models from its scope.
• It applies only to closed-source, frontier-level AI models, meaning models only their developers can fully inspect, that carry national security risks.
• AI companies get a 30-day window for the government to review a model before release, but compliance is entirely voluntary.
• The framework defines neither "state-of-the-art" nor "national security risk," leaving the terms that matter most open to interpretation.
• The White House has no plans to release the framework's details to the public.

The Trump administration has completed a framework for testing whether advanced AI systems pose cybersecurity dangers before they go public. But as first reported by Axios, it leaves out a large and growing part of the AI world before it even gets started.

What does the framework actually cover?

It covers only closed-source AI models with frontier capabilities that carry national security risks. A closed-source model is one where the underlying code and components are kept private, the way a recipe is locked in a vault. Open-source models, where anyone can download and inspect the inner workings, are excluded entirely.

The framework also explicitly states it can't be used to restrict an open-source model once it has been released. That's a notable carve-out, given that several of the most capable AI systems available today, including Meta's Llama family, are open-source. Our 5 August story previewed Tuesday's White House briefing before the framework was formally presented to Anthropic, OpenAI, and Google.

The meeting followed an executive order President Trump signed in June directing AI companies to share their most advanced models with the federal government before public release.

What does it mean for AI companies?

In practice, companies aren't legally required to comply. The guidelines are voluntary.

Under the framework, the government has 30 days to review a new model after a company submits it. That window applies only to closed-source models judged to be both frontier-level and a national security risk.

Neither of those conditions is defined anywhere in the document. For a policy built specifically to evaluate those things, that's a glaring gap. Frontier labs like OpenAI and Anthropic have reportedly been pushing for clearer guidance, hoping to understand where the line sits before they release something that draws government scrutiny. Smaller companies face the same problem: without clear definitions, it's hard to know whether a new model falls under the framework at all, or whether releasing it quietly could damage their standing with the administration.

The ambiguity also matters for open-weight models from overseas. Our 4 August story found that GLM-5.2, an open-weight model from China's Z.ai, came close to matching Western frontier systems on dangerous capabilities yet refused none of the harmful tasks it was given. A framework that can't touch open models won't reach that category of risk.

What happens next?

The administration isn't planning to publish the framework, which means outside experts and the public can't scrutinise the standards being applied to AI models before they reach everyday users.

That matters. AI models already shape medical records, hiring decisions, and customer service interactions affecting millions of people. A testing framework with undefined terms and no public accountability offers thin reassurance that those systems have been genuinely stress-tested.

The voluntary nature of the rules means a company that wants to move fast can simply choose not to participate.

Common questions

Does this affect AI tools I already use?

Not directly. The framework targets new models before release, not ones already on the market. Tools you use today are unaffected.

Why does it matter that open-source models are excluded?

Open-source AI models can be downloaded and modified by anyone, including people with harmful intentions. A framework that cannot address those models leaves a large portion of the risk picture untouched.

Is this a law?

No. The framework is voluntary guidance, not legislation. Companies face no legal penalty for ignoring it.