The U.S. government will now get an early look at the most powerful AI models before they are released to the public.
On May 5, 2026, the National Institute of Standards and Technology (NIST) announced agreements with Google, Microsoft, and xAI. These deals grant a special government task force pre-release access to test so-called 'frontier AI models' for national security risks. This isn't just a friendly review; the testing can happen in classified environments and even involve models with their safety features turned down or removed to see their full potential for misuse.
This development didn't happen overnight. It's the result of a clear trend of increasing government oversight. First, policies under the Biden administration established a norm of information sharing, requiring companies to report safety data on their models. Then, the Trump administration escalated this by implementing export controls on AI model weights in 2025, treating powerful AI less like software and more like a strategic asset that needs to be controlled.
Second, the government's own needs drove this change. The Pentagon recently finalized deals to use commercial AI for "all lawful purposes," including in classified operations. This created a strong demand for the government to independently verify that these tools are safe and reliable. A previous standoff with the AI company Anthropic, which resisted unrestricted use, highlighted the need for a standardized process to avoid future conflicts and ensure AI labs meet the government's terms.
Finally, the broader geopolitical context, particularly the U.S.-China tech rivalry, provides a compelling reason for this move. With Congress investigating risks from Chinese AI, ensuring domestic models are thoroughly vetted becomes a national priority. This structured testing program gives Washington greater bargaining power, aligning corporate AI development with national security interests. For companies like Google and Microsoft, passing these tests could become a golden ticket to massive government contracts, potentially adding billions to their revenue.
- Frontier AI Models: Refers to the most powerful and capable AI models currently in existence, which have the potential for both significant benefits and substantial risks.
- Model Weights: These are the numerical parameters of an AI model that are learned during the training process. They essentially contain the 'knowledge' the model has acquired and are crucial to its functionality.
- Red-Teaming: A security testing method where a dedicated team acts as an adversary to find and exploit vulnerabilities in a system. In AI, this involves trying to 'jailbreak' the model or make it produce harmful outputs.