TL;DR: Microsoft, Google DeepMind, and xAI have signed agreements to allow the U.S. government to test their frontier AI models for national security risks before public release. The move signals a major shift by the Trump administration from a hands-off approach to active oversight of advanced AI systems.
The U.S. Government Steps Inside the AI Black Box
In a landmark pivot for AI regulation, the U.S. federal government has secured unprecedented pre-deployment access to the world’s most advanced artificial intelligence models. The Department of Commerce’s Center for AI Standards and Innovation (CAISI) announced agreements with Microsoft, Google DeepMind, and xAI to evaluate their frontier models for national security risks before they reach the public.
The agreements grant government scientists the ability to probe these systems for unexpected behaviors, vulnerabilities, and potential misuse, ranging from cyberattacks to military applications. To conduct rigorous testing, developers will frequently provide CAISI with “jailbroken” versions of their models—stripped of standard safety guardrails—allowing evaluators from across the government, including the TRAINS Taskforce, to thoroughly assess their raw capabilities in classified environments.
Anthropic’s Mythos and the Catalyst for Oversight
The urgency behind these agreements stems from growing alarm in Washington over the capabilities of next-generation AI. The recent unveiling of Anthropic’s Mythos model has been a primary catalyst. Mythos demonstrated unprecedented proficiency in cybersecurity tests, raising fears that such tools could supercharge malicious actors if released without rigorous vetting.
Notably absent from the CAISI agreements is Anthropic itself. The company has been embroiled in a public dispute and legal battle with the Trump administration over the ethics and safety of deploying AI in military and national security contexts. This absence highlights a growing fracture within Silicon Valley regarding collaboration with the defense and intelligence apparatus.
Background
The Center for AI Standards and Innovation (CAISI), under the Department of Commerce’s NIST, is the U.S. government’s primary hub for AI model testing. Initially the U.S. Artificial Intelligence Safety Institute under Biden, its early focus was on voluntary safety standards and protocols. Elizabeth Kelly, a tech adviser, led the institute before transitioning to Anthropic.
The Trump administration, initially noninterventionist, recalibrated its stance due to escalating frontier AI capabilities and potential national security threats. New agreements now mandate active vetting of AI systems, marking a definitive shift toward proactive government oversight.
