The U.S. government just got a key to the back door of the biggest AI labs - five of them, actually. Google, Microsoft, and xAI agreed to hand their models over to a federal review team before public launch, joining OpenAI and Anthropic on a list that now covers most of the frontier.
The reviewer is the Center for AI Standards and Innovation, or CAISI, which sits inside the Commerce Department.
Its job is to test AI models before release for security risks, including national security threats. CAISI has already finished more than 40 evaluations, and some of those models are not yet public.
The labs hand over versions of their models with safeguards reduced or turned off, so the team can probe for weaknesses an attacker might exploit. That trust runs in both directions, since the labs have to expose unfinished work and the government has to keep it confidential.
Anthropic's Mythos model is the trigger, since the system reportedly turned out to be very good at finding weak spots in cybersecurity defenses. That set off alarms inside the White House. Chief of Staff Susie Wiles, Treasury Secretary Scott Bessent, and National Cyber Director Sean Cairncross are all involved in the response, and the administration has even pushed back on Anthropic's plan to widen access to Mythos.
The bigger picture: President Trump's AI Action Plan, released in July, made CAISI the lead on national security AI evaluations. The deals announced this week put real labs against that mandate, which is what the plan needed to actually work.
CAISI's work is happening as the Defense Department fights Anthropic in court over whether the Pentagon can label the company a supply chain risk.
Defense Secretary Pete Hegseth and President Trump have set out a six-month timeline to phase out Anthropic tools across the federal government, and a White House memo on AI use is expected soon.
The team running CAISI has had its own drama, since Director Chris Fall took the job after Collin Burns, a former Anthropic researcher, was forced out within days of starting.
Whether CAISI's reviews become formal regulation. The center is not yet written into law, though there is bipartisan legislation in the Senate to change that. Once a federal agency has a habit of testing models early, it gets harder to take that role away.
