← Back to feed

The Trump administration is considering a pre-release government review process for major AI models, according to reporting from the New York Times confirmed by Reuters. The proposed review would not block releases outright but would give federal agencies early access to test advanced models before the public receives them.

This is a significant reversal. The same administration rolled back Biden-era reporting requirements for powerful AI systems in early 2025, arguing that heavy-handed regulation would slow US progress against China.

What Changed

Anthropic’s Mythos model changed the calculus. Mythos is the model Anthropic describes as unusually capable at discovering and chaining software security flaws — a capability the company deemed too risky for broad release. The model is currently accessible to fewer than 50 organizations after the White House blocked Anthropic’s push to expand to 120.

The concern is specific: a model capable of scanning codebases, identifying weak points, and generating attack paths faster than a human security team represents a different category of risk from a general-purpose assistant. Bank infrastructure, government systems, cloud providers, and critical infrastructure all present potential targets if that capability reaches adversarial actors without restriction.

Scope of the Proposed Review

The review process, if implemented, would focus on cyber capability — specifically whether a model can discover, exploit, or chain software vulnerabilities at a scale that materially shifts real-world security risk. The review would not necessarily be a gatekeeping mechanism; the framing is closer to early access for federal evaluation.

No legislation has been introduced. The discussions are internal to the administration. The New York Times characterized the conversations as early-stage.

The Policy Tension

The administration is running competing imperatives. On one side: the view that US AI leadership requires moving fast and that regulatory overhead cedes ground to Chinese competitors. On the other: intelligence agency and cybersecurity community pressure that frontier models with offensive cyber applications require some form of national security review before broad deployment.

The Anthropic-NSA dynamic is illustrative. NSA deployed Mythos for offensive and defensive cyber work under a classified arrangement. The same White House that approved that deployment is now considering whether similar models should face a pre-release review before public access — not before government access.

Industry Implication

If formalized, pre-release government review would add lead time and unpredictability to model launches for any lab operating at the frontier. It would also create a two-tier market: models cleared for government use (potentially ahead of public release) and models pending civilian deployment. Labs with existing government relationships — Anthropic, OpenAI, Google — would have structural advantages in navigating that process. Labs without would face additional friction.

The policy has not been adopted. But that it is being discussed at all marks a meaningful shift from the administration’s starting position.