The US Commerce Department removed details from its website about its May 5 agreement with Google, xAI, and Microsoft to test their AI models for security flaws
Days earlier, Commerce’s CAISI said Google, Microsoft and xAI would join OpenAI and Anthropic in giving the government early access to assess models before public release. Removing the published details does not establish that the testing arrangement ended, but it makes its terms and status less visible.
The change follows separate Commerce action to withdraw a proposed tightening of AI-chip export rules, underscoring that the department’s AI policy posture is being revised across both model oversight and compute controls.
First-order effects
Google, Microsoft and xAI lose a public reference point for their stated security-testing commitments with Commerce; outside stakeholders have less visibility into the scope of the arrangement.
Commerce retains the ability to engage frontier-model developers, but its public accountability for how those evaluations are structured is reduced unless it provides details through another channel.
Second-order effects
OpenAI and Anthropic’s earlier-access arrangements become harder to compare with those involving the newly added companies, limiting a common public benchmark for voluntary pre-release evaluation.
Other model developers may face less reputational pressure to disclose the terms of government safety testing when leading participants’ agreements are no longer documented publicly.
Third-order effects
If government model evaluations increasingly operate through unpublished, company-specific arrangements, AI oversight could shift toward private coordination rather than transparent, standardized safety processes.
The combination of less-visible model-testing agreements and a withdrawn chip-export proposal points to a more fluid federal AI-governance approach, with the durability of formal rules and public commitments uncertain.
The trend: This is a data point in the shift from publicly articulated AI guardrails toward more discretionary, bilateral government engagement with major model providers.
The website was removed because of sensitivities within the White House Office of the National Cyber Director, amid an administration turf war over AI evaluations https://www.washingtonpost.com/ ...
That @NIST CAISI announcement from last week about Google DeepMind, xAI and Microsoft signing new deals for pre-deployment testing is gone from their website, link no longer found, and I can't get anyone from Commerce to tell me why. See for yourself: https://www.nist.gov/...
Worrying zero-sum mindsets here, if this coverage is accurate. It's pretty obvious both CAISI and the intelligence community have important but different roles to play https://x.com/...
Thanks to @Cat_Zakrzewski, @nakashimae, and @nitashatiku for including my thoughts in their piece in @washingtonpost on how the administration is reevaluating its approach to AI policy in light of recent capability developments. [image]
The Washington Post is reporting that the U.S. intelligence community is interested in a much more prominent role in reviewing AI models. The intelligence community has a complicated history with powerful new technologies. This is the community that in the 1990s opposed strong
@Cat_Zakrzewski ... “Biden established the US AI Safety Institute under the Commerce [Department]... In a sign of the new administration's shift, Trump removed “Safety” from the agency's name & rebranded it as the “Center for AI Standards & Innovation.” A rose by any other name, …
NEW: The Trump administration is sharply split over a proposal to give spy agencies a bigger role in evaluating AI models. One person described it as a ‘knife fight.’ w/ @nakashimae @nitashatiku https://www.washingtonpost.com/ ...