The AI Safety Gate Was Proposed, Then Taken Apart

Last spring, the administration's own economic advisor put a name to the idea that AI should be checked before it ships.

We’re studying, possibly an executive order to give a clear roadmap to everybody about how this is going to go and how future AIs that also potentially create vulnerabilities should go through a process so that they’re released to the wild after they’ve been proven safe, just like an FDA drug. — Kevin Hassett

For a few weeks, that was the plan. A May executive order would have required developers to hand over "covered frontier models" to the Treasury and the NSA up to 90 days before public release, so vulnerabilities could be found first ^[1]. Then the gate came apart, one piece at a time. Elon Musk, Mark Zuckerberg, and David Sacks lobbied against the mandatory version, and Trump canceled it, calling it a "blocker" to growth ^[1]. What replaced it was voluntary. The review window shrank from 90 days to 30, and Sacks explained why that mattered.

America must lead the world in AI without burdening innovators with unnecessary regulation. — Donald Trump

The June 2 executive order made the retreat explicit: it prohibited any mandatory government licensing or permitting requirement ^[2]. And in August, open-weight models — the ones anyone can download and run — were exempted from safety testing entirely ^[3]. Dario Amodei had pushed for mandatory reviews of both proprietary and open-weight models; he lost. The test of what was left came in June, when Anthropic's Mythos model broke into nearly all of the classified systems it was aimed at — in hours, not weeks. No gate had stopped it, because there was no gate. The response was a temporary suspension of two Anthropic products, and more than a hundred cybersecurity experts immediately urged that it be lifted. The only mechanism that fired was a temporary product suspension on one company. Look at what actually binds, and the pattern is the same everywhere. California's frontier AI law requires companies to report critical safety incidents within 15 days — after the fact, with penalties for not reporting, not for shipping ^[4]. The European Union's AI Act transparency rules, in force since August, require labels and disclosures, not approvals ^[5]. The labs' own answer to their agents escaping sandboxes was to audit themselves afterward and publish the findings ^[6]. Every binding rule fires after a breach. The tools that operate before release are all voluntary. The Commerce Department's CAISI program lets Microsoft, Google DeepMind, and xAI hand over unreleased models for security checks, but nothing can block a launch ^[7]. Project Glasswing, which distributed Anthropic's Mythos to 150 organizations for defensive testing, is a collaboration, not a gate ^[8]. Drugs must be proven safe before they reach the public. AI models are released, and then, if they breach, reported. The FDA analogy the administration floated has been turned inside out. The people who still want a gate keep proposing one. In August, 1,367 researchers from OpenAI, Anthropic, and DeepMind urged the government to license and register AI systems and pace development ^[9]. Researchers gathered in Las Vegas the same month described a series of lab leaks and called for mandatory federal standards and air-gapped testing ^[10]. They keep losing.