Google DeepMind, Microsoft, and Elon Musk’s xAI have struck deals with the U.S. Commerce Department. They will hand over early versions of their most advanced AI models for government scrutiny before public launch. The agreements, announced Tuesday by the Center for AI Standards and Innovation—or CAISI—mark a sharp turn. All five major frontier AI developers now fall under this voluntary pre-deployment review net. OpenAI and Anthropic paved the way back in 2024. NIST press release details how CAISI has already run more than 40 evaluations, some on still-unreleased state-of-the-art systems.
CAISI Director Chris Fall put it bluntly. “Independent, rigorous measurement science is essential to understanding frontier AI and its national security implications,” he said. “These expanded industry collaborations help us scale our work in the public interest at a critical moment.” The partnerships got a refresh. They now align with directives from Commerce Secretary Howard Lutnick and President Donald Trump’s AI Action Plan. Developers often strip safeguards from test models. That lets evaluators probe national security risks in classified settings. The TRAINS Taskforce—interagency experts—provides feedback to sharpen voluntary fixes.
But. This didn’t happen in a vacuum. Anthropic’s Claude Mythos lit the fuse. The model spots flaws in every major operating system and browser, Anthropic warned. Too risky for open release. Government agencies tapped it anyway to hunt vulnerabilities in federal software. Yet a February contract spat with the Pentagon blew up. Anthropic balked at unrestricted access demands. Defense Secretary Pete Hegseth slapped a supply-chain risk label on the firm. Military branches and contractors got orders to ditch its tech within six months. Anthropic sued. A federal appeals court denied a pause. Forbes ties the timing to this rift, noting Trump’s earlier hands-off vibe—’We have to grow that baby and let that baby thrive’—has flipped.
Mythos fears rippled wide. White House officials huddled with tech execs from Anthropic, Google, and OpenAI. Discussions centered on AI-fueled cyberattacks. The New York Times reported Monday that Trump eyes an executive order. It would forge a working group of industry leaders and officials to vet powerful models upfront. No blocking power, just first looks. Similar to U.K. safety checks. Axios detailed White House cyber director meetings last week with tech firms and trade groups. Mythos prompted calls for Pentagon safety tests on models hitting federal networks. The New York Times flagged White House worries over political fallout from AI hacks.
Pentagon moves amplified the pressure. Last week, it inked pacts with OpenAI, Alphabet’s Google, Nvidia, SpaceX (via xAI ties), Microsoft, Amazon, and Reflection AI. Their tools now run on classified IL6 and IL7 networks for ‘lawful operational use.’ Data crunching. Situational awareness. Decision aids for warfighters. Google faced employee pushback—nearly 600 signed a letter against military AI ties, echoing 2018’s Project Maven exit. But leaders held firm. Rivals like OpenAI and xAI signed on too. Anthropic’s holdout cost it dearly. The Verge notes the voluntary nature: no legal teeth to halt releases.
xAI’s entry stands out. Elon Musk’s outfit, tied to SpaceX, joins despite Musk’s past gripes on regulation. Bloomberg reported the firms joining OpenAI and Anthropic in pre-release handovers. Bloomberg highlighted renegotiations for Trump’s plan. The Wall Street Journal confirmed early model shares for capability and security checks. WSJ. Financial stakes? Pentagon deals run into hundreds of millions, like Anthropic’s failed $200 million bid.
Industry insiders see a new normal. Five labs dominate frontier AI. All now vetted by one understaffed office—under 200 people, no statutory clout. Yet it works. Post-deployment checks follow. Info flows back to firms for tweaks. International rivalry looms. China watches. U.S. aims to stay ahead without stifling growth. Trump tapped an AI advisory panel in March: Zuckerberg, Ellison, Huang, Dell. They shape policy on science and tech.
Critics question the setup. Voluntary only goes so far. What if a firm bolts, Mythos-style? Executive order could formalize it. White House cyber fears—hacks exploiting AI—drive urgency. Anthropic cofounder Jack Clark said talks continue on Mythos despite bad blood. Federal eyes seek its cyber-hunting edge.
So where next? CAISI scales up. More models. Deeper classified tests. Trump’s order, if inked, binds execs closer. Big Tech bends. National security trumps pure speed. AI’s dual-use edge—civilian boon, military game—demands balance. Developers comply. For now.
WebProNews is an iEntry Publication