The Trump administration is asking Meta to let government officials review its Llama artificial intelligence models for security problems. The unusual part: federal agencies are already using these AI models in their daily work.

Meta's Llama models were approved for government use in September 2025. Since then, federal agencies have been using them to review contracts, fix IT problems, and handle national security tasks. Meta has confirmed that the models help identify safe places for aircraft to land and translate languages — jobs where the AI needs to be accurate and trustworthy.

The timing is backwards in an important way. Normally you check something for safety before you use it. This time, the government used the AI first and is asking to review it now. That matters because Llama is "open-weight," meaning anyone can take the model, modify it, and use their own version. A government security review of Meta's original model would not catch all the different versions people have created and are using elsewhere.

Meta has been cooperating closely with the Trump administration. The company announced it would invest $600 billion in the U.S. through 2028, following a September 2025 meeting with major tech leaders. It also pledged over $20 million toward AI education. That partnership is why the government asked for a voluntary review rather than ordering one — both sides benefit from appearing to cooperate.

There is a catch to calling this "voluntary." The government gets to look at information about the AI and influence how it is used, without creating legal problems for Meta. Meta gets to show it is cooperating while keeping some control over what it shares. This works fine until something goes wrong. If a government-run copy of Llama causes a security incident, the word "voluntary" becomes a problem, not a protection.

No one has said yet exactly what the review would check. It would need to cover whether the AI can be tricked into doing harmful things, where the training data came from, and how well it resists being manipulated through clever prompts — especially important when the AI makes decisions on its own, like in the aircraft-landing example. Whether the government has the expertise to evaluate models this complex is unclear.

For government and business AI use overall, this sets a pattern to watch. The U.S. government has historically been slow to create rules that keep up with how fast AI advances. NIST's AI Risk Management Framework gives useful guidance, but it is not a tool for real-time checking. If this review creates a repeatable process, it could become the template for how all federal agencies buy AI. If it fails, it joins a long list of AI governance efforts that created paperwork without creating actual standards.

Meta has not said publicly whether it will participate. Given how much federal work Meta already does with AI, refusing would be unusual. The real discussion will probably be about what Meta has to share — whether the government can see the model's weights, training notes, and security test reports.

Open-weight AI models are now at a stage where these questions cannot be avoided. Llama is not the only open model being used in important government work, but it is the most visible. How this plays out will affect every other AI company trying to sell models to the federal government.