From Mocking Safety Tests to Rebuilding Them

The Trump administration has sharply changed its position on frontier AI oversight, signing new agreements with Google DeepMind, Microsoft, and xAI to conduct government safety checks on advanced models before and after release. The shift is notable not only for what it does, but for how directly it departs from the administration’s earlier stance.

According to Ars Technica, President Donald Trump had previously dismissed voluntary AI safety checks associated with the Biden era as regulatory overreach that hindered innovation. Soon after taking office, the administration even rebranded the US AI Safety Institute as the Center for AI Standards and Innovation, or CAISI, removing the word “safety” from the title in what the source describes as a pointed jab at Joe Biden.

Now that same administration is restoring a version of the policy framework it had derided.

The Mythos Trigger

The apparent turning point was Anthropic’s decision not to release its latest Claude Mythos model. Ars Technica reports that the company judged the model too risky to publish because bad actors might exploit its advanced cybersecurity capabilities. That decision appears to have had a strong effect inside Washington.

White House National Economic Council Director Kevin Hassett told Fortune, according to the source text, that Trump may soon issue an executive order requiring government testing of advanced AI systems before release. If that happens, the administration would move from renewed voluntary cooperation to a formal pre-release evaluation regime.

The political irony is difficult to miss. A White House that had positioned safety oversight as a brake on progress is now considering a mandate after a frontier model raised concerns serious enough that its developer withheld it.

CAISI Steps Into a Larger Role

CAISI’s own announcement explicitly said the new agreements “build on” Biden’s earlier policy. That language matters because it acknowledges continuity even after a period of rhetorical rupture. The administration may have changed the branding, but the operational need for structured model evaluation appears to have reasserted itself.

CAISI Director Chris Fall said the expanded industry collaborations would help the center scale its work “in the public interest at a critical moment.” The source text says CAISI has completed around 40 evaluations so far, including tests on frontier models that had not yet been released.

Those evaluations are especially consequential because CAISI often receives access to systems with reduced or removed safeguards. According to the agency, that allows evaluators to assess national-security-related capabilities and risks more thoroughly than they could through public-facing deployments alone.

In practice, that means the government is not only reviewing polished products after launch. It is examining what models can do when some safety layers are stripped back, a much more probing form of capability assessment.

National Security Is Driving the Policy Shift

The administration’s new posture is framed heavily around national security. CAISI says an interagency group of experts has formed a task force focused on AI national security concerns, intended to keep evaluators aligned with emerging risks across government.

This emphasis is important because it helps explain the political repositioning. General debates about AI ethics or consumer harms can divide policymakers along familiar ideological lines. National security questions often rearrange those lines more quickly. A model thought capable of dangerous cyber misuse is easier to cast as a strategic threat than as a conventional technology-policy dispute.

That does not settle the larger argument over how AI should be governed, but it clarifies what is currently moving the administration. The concern is not abstract safety branding. It is the possibility that frontier systems could create operational advantages for hostile actors before government evaluators understand the risk profile.

Why This Reversal Matters

The significance of the new agreements lies in both policy and signal. On policy, they restore a channel through which major AI companies submit frontier systems to government scrutiny before and after release. On signal, they indicate that even an administration rhetorically hostile to “AI safety” can return to structured oversight when capabilities rise fast enough.

The participating firms also matter. Google DeepMind, Microsoft, and xAI are among the companies building or backing leading-edge systems, so their involvement gives the framework practical weight. If a future executive order makes testing mandatory, the government will already have active relationships and some operational history to build on.

At the same time, the current shift leaves open several unresolved questions. The source text notes that some firms have signed agreements, but it does not provide a complete roster or enforcement details. Nor does it explain what thresholds would define an “advanced” system for mandatory testing if Trump proceeds with an executive order.

Even so, the direction of travel is clear. A White House that had tried to downplay AI safety has been pulled back toward it by the behavior of frontier models and the warnings of the companies building them.

An AI Policy Debate Reframed by Capability

The episode is a reminder that AI policy arguments can change quickly when theoretical concerns become operational ones. The Trump administration did not gradually evolve toward the Biden view out of rhetorical compromise. It appears to have reacted to a concrete case in which a major developer concluded a model was too risky to release.

That distinction matters. In the emerging AI landscape, capability shocks may shape policy faster than ideological consistency does. And if the administration follows through with mandatory pre-release testing, one of the strongest recent US moves on frontier AI evaluation will have emerged from a president who initially treated such oversight as unnecessary.

  • The Trump administration signed new voluntary AI safety testing agreements with Google DeepMind, Microsoft, and xAI.
  • The move reverses its earlier dismissal of Biden-era AI safety checks.
  • Anthropic’s decision not to release Claude Mythos appears to have accelerated the shift.
  • The White House may next pursue an executive order mandating pre-release testing of advanced AI systems.

This article is based on reporting by Ars Technica. Read the original article.