Issue 353

The U.S. government has undergone a sharp shift in its stance on AI regulation. NIST announced the formation of the interagency working group TRAINS (Testing Risks of AI for National Security), led by the Commerce Department's Center for AI Standards and Innovation (CAISI), in collaboration with multiple federal agencies including the Department of Defense, Department of Energy, Department of Homeland Security, National Security Agency, and National Institutes of Health, to conduct pre-release national security risk assessments of frontier AI models. The White House is also considering an executive order requiring AI models to obtain approval before deployment.

This policy stands in stark contrast to the Trump administration's previous position. After taking office in January 2025, President Trump assigned three key advisors to develop an AI action plan aimed at 'maintaining and enhancing U.S. global AI dominance' by suspending or repealing Biden-era regulations. In 2023, the Biden administration had issued an executive order requiring developers to notify the government when training models at approximately the 1 trillion parameter scale.

The direct catalyst for the policy shift was Anthropic's announcement of the Claude Mythos Preview model. The model can autonomously exploit vulnerabilities in mainstream operating systems and applications, and Anthropic had shared it with 50 organizations for vulnerability detection and patching. However, the White House last week opposed the company's plan to expand the preview to an additional 70 organizations, citing national security concerns and questioning whether Anthropic had sufficient computing power to simultaneously serve existing users and the government.

Earlier, in March 2026, Anthropic had attempted to restrict Claude's use in military surveillance and autonomous weapons, but the White House rejected any restrictions and completely banned the model for military use. A month later, Anthropic announced the capabilities of Mythos Preview, drawing government attention.

TRAINS will focus on assessing models for demonstrable risks in cybersecurity, biosecurity, and chemical weapons. NIST has not disclosed the specific benchmarks for evaluation, but CAISI previously used nine widely used public benchmarks (covering cybersecurity, programming, mathematics, natural sciences, and abstract reasoning) along with an internal test, PortBench (cross-language porting of command-line tools), to rank the capabilities of models such as DeepSeek V4 Pro.

Google, Microsoft, and xAI have agreed to provide models with limited or no guardrails for evaluation; Anthropic and OpenAI had already agreed to similar terms in 2024. These agreements aim to foster public-private collaboration in studying capability assessment, risk evaluation, and mitigation measures.