The same administration that ripped up AI safety rules six months ago is now building a federal approval process for new models.
The Summary
- The White House is drafting an executive order to create a vetting system for AI models before release, targeting cyber risks to business and government networks
- This represents a reversal of Trump's earlier stance, with officials forming an AI working group to discuss oversight and review processes
- The move would give the federal government pre-release approval power over models like Anthropic's Mythos, inserting Washington into the launch timeline of every major AI lab
The Signal
The proposed vetting system targets a real problem. AI models are increasingly used to probe networks and automate cyberattacks, and government systems remain easy targets. If a hostile actor gets early access to a capable model, they can weaponize it before defenses adapt. Pre-release review could catch models with obvious exploit potential before they hit API endpoints.
But the mechanism matters more than the intent. A vetting system run by career security professionals looking for specific technical vulnerabilities is different from a regulatory chokepoint that lets political appointees slow-walk approvals. The order's language will reveal which version we're getting.
"The gap between 'security review' and 'permission slip' is where innovation goes to die."
The timing is the tell. Trump's administration previously opposed AI safety regulations, dismantling Biden-era frameworks. Now they're building a pre-release approval process, which is a heavier intervention than anything Biden proposed. Three explanations fit:
- Mounting pressure from defense and intelligence agencies who see AI-enabled intrusions accelerating
- A play to assert federal control over AI development before states or Congress act first
- Recognition that "move fast and break things" doesn't work when the things breaking are power grids and financial systems
The AI working group signals this is still in formation. Working groups are where policy goes to get workshopped, watered down, or weaponized depending on who's in the room. If the group includes technical experts from NIST and CISA, expect a narrow, threat-focused vetting process. If it's heavy with political operatives and industry lobbyists, expect a tool that favors incumbents.
For AI labs, this creates a new chokepoint. Right now, a company can train a model and launch it the same week. Under a vetting regime, launch dates depend on federal review timelines. That's leverage. It's also a moat for companies that can afford compliance teams and lobbyists.
The Implication
Watch the working group membership and the draft order's technical requirements. If vetting focuses on narrow cybersecurity tests with clear pass/fail criteria, it's workable. If it's vague standards with discretionary approval, it's regulatory capture waiting to happen.
For builders, this is a reminder that the window for "permission-optional" AI development is closing. If you're launching agents or models that touch critical infrastructure, government systems, or financial networks, expect scrutiny. Build your security documentation now, not when you're in a review queue.