In December, Trump told AI companies they were "free to innovate without cumbersome regulation." Now his administration is weighing whether those same companies need government approval before releasing their most powerful models.
The reason is concrete enough. As we covered over the weekend, Anthropic's unreleased Mythos model can autonomously discover software vulnerabilities that human researchers missed for decades. Then the UK's AI Safety Institute evaluated OpenAI's GPT-5.5 and found the model solved a reverse-engineering challenge in about 10 minutes for $1.73 in API costs. A human expert with professional tools took roughly 12 hours. A year ago, no AI could do any of this.
Two different companies crossed the same dangerous capability threshold within weeks. That's what turned this from one company's problem into a policy question.
The administration moved quickly. On Tuesday, the federal government's AI testing body, CAISI, announced that Microsoft, Google, and xAI had all signed agreements for pre-release national security testing of their frontier models. Anthropic and OpenAI already had similar deals in place. CAISI has completed more than 40 pre-deployment evaluations to date, including on models that haven't been released to the public.
Making those agreements mandatory is a different question entirely. Eli Dourado, head of strategic investments at the Astera Institute, argued in a public post that mandatory model reviews would conflict with First Amendment protections against prior restraint. A RAND Corporation research report warned that the scarcity of real-world examples in this nascent field should not lead to a "false sense of security" regarding AI infrastructure vulnerabilities.
IBM CEO Arvind Krishna offered conditional support, saying government review works as long as it stays fast, within "a few days or a few weeks." Anything beyond that, he said, "would not be good."
This is the same administration that revoked Biden's AI safety executive order on its first day, with allies like Senator Ted Cruz calling that framework "barriers to innovation disguised as safety measures." Now it's quietly rebuilding a version of what it tore down, framed around national security rather than safety.
When the most deregulatory administration in recent memory starts building AI oversight, it tells you something about where the technology actually is. The models scared them. The real question isn't whether the government should be involved, because that debate appears to be over. It's whether what they build will be fast enough to keep pace with a release cycle measured in weeks, or slow enough that companies route around it entirely. A model that does 12 hours of expert work in 10 minutes doesn't wait for interagency review.
