US Commerce deletes security-test details

The fight here is over one phrase — “pre-deployment evaluations.” That sounds dry, but in AI policy it means something big: government experts getting a look at powerful models before the public does. Last week, the Commerce Department’s AI office said Google DeepMind, Microsoft, and xAI had agreed to exactly that. Then the detailed announcement disappeared from the web. The result is a very Washington kind of mess — the government signaled tougher oversight, industry and policy watchers reacted, and now everyone is left parsing what was real policy and what was just over-eager wording. ### What was actually posted? On May 5, Commerce’s Center for AI Standards and Innovation — CAISI, inside NIST — announced new agreements with Google DeepMind, Microsoft, and xAI. The text said CAISI would conduct “pre-deployment evaluations and targeted research” on frontier AI systems, and reporting at the time said the models would be tested in classified environments. That was a bigger claim than a normal standards partnership. It implied the U.S. government would have access before release, not just after problems showed up. (nextgov.com) ### Why did that wording matter so much? Because “before release” is the line the AI industry has spent years resisting. Voluntary safety work is one thing. A standing government lane into unreleased models is another. Even if the agreements were technically voluntary, the public message landed like a pre-launch review regime for frontier AI. That is why the announcement got so much attention so quickly. (nextgov.com) ### Was this brand new policy? Not exactly. CAISI already had earlier agreements with OpenAI and Anthropic, signed in August 2024 when the office was still called the U.S. AI Safety Institute. Those deals also allowed access to major new models before and after public release for research, testing, and evaluation. The new May 5 agreements extended that network to Google DeepMind, Microsoft, and xAI — basically bringing all five major U.S. frontier labs into the same voluntary testing orbit. (bloomberg.com) ### So what changed under Trump? The branding and the politics. In June 2025, Commerce reworked the Biden-era AI Safety Institute into CAISI and framed it as more innovation-friendly and less regulatory. NIST’s current CAISI page still says the office will lead evaluations and voluntary agreements, but it also stresses support for industry, standards, and protecting U.S. firms from what it calls burdensome foreign regulation. That makes a hard-sounding “pre-deployment” review message feel out of step with the administration’s own pitch. (nist.gov) ### Why delete the details now? Commerce has not publicly explained the removal. But the obvious read is that the wording ran ahead of the politics. If you are trying to sell CAISI as a light-touch partner, posting language that sounds like federal gatekeeping is a problem. The catch is that the deletion does not erase the broader fact that these agreements were announced and widely described as giving CAISI early model access. That leaves the substance murky, not settled. (nextgov.com) ### Does this mean the testing is gone? Not necessarily. CAISI’s public mission still includes voluntary agreements with private AI developers and evaluations tied to national-security risks like cybersecurity, biosecurity, and chemical weapons. What vanished was the clean public statement spelling out the Google DeepMind, Microsoft, and xAI arrangement in those terms. So the most likely takeaway is not “the program died.” It is “the administration got nervous about how explicit it looked.” That last part is an inference, but it fits the remaining public record. (thenextweb.com) ### Why should anyone outside Washington care? Because frontier AI companies need stable rules, and right now they are getting policy vibes instead. One week the government sounds like a pre-release reviewer. The next week the page is gone. That kind of whiplash matters for labs deciding what to disclose, investors trying to price regulatory risk, and allies wondering whether the U.S. wants real oversight or just the appearance of it. (nist.gov) ### Bottom line? The deletion is small, but the signal is not. Washington is still trying to have it both ways — tougher scrutiny for powerful AI, but without sounding like it is regulating the labs too directly. That balancing act keeps breaking in public. (nextgov.com)