The U.S. government’s AI security center will evaluate frontier models from Google, Microsoft and xAI before their release to determine whether the models’ advanced capabilities pose cybersecurity risks.
The newly announced plan for the National Institute of Standards and Technology’s (NIST) Center for AI Standards and Innovation (CAISI) to conduct “pre-deployment evaluations” represents the U.S. government’s most significant attempt yet to get ahead of security threats from powerful AI systems.
“Independent, rigorous measurement science is essential to understanding frontier AI and its national security implications,” CAISI Director Chris Fall said in a statement. “These expanded industry collaborations help us scale our work in the public interest at a critical moment.”
NIST said the partnerships would help the agency and the tech companies exchange information, spur “voluntary product improvements” and ensure the government had a “clear understanding” of what AI models were capable of doing. An interagency task force at CAISI will allow officials from across the government to test the models, including in classified settings.
Natasha Crampton, Microsoft’s chief responsible AI officer, said in a LinkedIn post that tech companies couldn’t conduct “evaluations tied to national security and public safety” on their own.
“They require close collaboration between industry and governments with deep technical and security expertise,” she wrote, adding that Microsoft will apply what it learns “directly into how we design, test, and deploy AI — and share best practices to help strengthen AI testing more broadly.”
The arrangement represents a significant reversal for the Trump administration, which previously eliminated AI security review measures that it called overly burdensome.
The White House began rethinking its hands-off approach to AI after Anthropic announced that its latest model, Claude Mythos, was too dangerous to publicly release, because of its alarming ability to find serious software vulnerabilities. In addition to the new voluntary CAISI evaluations, the Trump administration is also considering instituting mandatory government reviews of all new AI models.
It remains unclear what testing standards CAISI will use for its evaluations. The outcomes that the agency defines as trustworthy and secure could be difficult to establish, according to Devin Lynch, a former director for cyber policy and strategy implementation at the White House Office of the National Cyber Director.
“Capability assessments are only as good as the threat models behind them,” Lynch wrote on LinkedIn. “CAISI will need to define, and publish, what it’s testing for, not just who it’s testing with.”
