China Thwarts Meta’s Agentic Ambition, U.S. Evaluates Upcoming Models, AI Diagnoses Mammograms

The National Institute of Standards and Technology just flipped a switch that had been set to 'stay out of the way' since January. A new group called TRAINS, run out of NIST's Center for AI Standards and Innovation, will check upcoming AI models for cybersecurity, biosecurity, and chemical-weapons risks before they ever reach the public. Google, Microsoft, and xAI have agreed to hand over versions of their models with weaker or no guardrails so testers can see what the systems can actually do; Anthropic and OpenAI signed similar deals back in 2024.

What makes TRAINS different from other NIST efforts is who's in the room. It pulls in the Departments of Commerce, Defense, Energy, and Homeland Security, plus the National Security Administration and the National Institutes of Health, and it's built to move fast rather than run the usual slow bureaucratic evaluation cycle. NIST hasn't said which benchmarks TRAINS will actually use, though it has pointed to an earlier CAISI comparison of DeepSeek V4 Pro against rivals using nine public benchmarks and an internal test called PortBench.

The timing isn't random. A month ago Anthropic raised alarms in government circles when it revealed that its unreleased Claude Mythos Preview model could find and exploit vulnerabilities in widely used software. That model has already been shared with 50 organizations using it to patch their own systems, and Anthropic wanted to expand access to 70 more. Last week the White House said no, citing national security worries and doubts about whether Anthropic has enough computing power to serve both those users and government needs. Anthropic hasn't said whether it will push back on that decision.

There's a bigger arc here too. Earlier this year Anthropic tried to restrict military use of Claude for surveillance and autonomous weapons, and the administration responded by banning the model from military use altogether rather than accepting limits. Combine that with the new pre-release testing regime and an executive order reportedly under consideration that would make model approval mandatory, and the picture is one of a government that spent its first year in office tearing down Biden-era AI rules, only to start rebuilding a different set of guardrails once models started showing they could do real damage.

None of the companies involved are required to submit models yet — these are voluntary agreements for now. But if that executive order lands, voluntary becomes mandatory, and the shape of who gets to release what, and when, starts running through Washington instead of Silicon Valley boardrooms.