Trump and Pre-Launch AI Controls
TL;DR: The Trump administration has reversed its position on AI: after dismantling Biden's rules, it's now studying new safety rules for models before launch. Microsoft, Google DeepMind, and xAI have already voluntarily agreed to CAISI reviews, the US Department of Commerce body. The turning point is triggered by Anthropic's Mythos model.
The administration that dismantled Biden's AI rules on its first day in office is now building its own oversight system. The direction is the same — government oversight before launch — but the path is different: voluntary agreements with big tech instead of imposed regulation.
What CAISI Is and What It Does
CAISI — the Center for AI Standards and Innovation — is a body within the US Department of Commerce. Its role is to conduct technical evaluations of frontier AI models before they're made available to the public.
It's not a censorship body: it evaluates the capabilities, safety risks, and potential for malicious use of the most powerful models. The result is a technical report that informs national security decisions and industrial policy.
On May 5, 2026, CAISI announced formal agreements with Google DeepMind, Microsoft, and xAI for pre-launch review of their respective models. OpenAI and Anthropic already had similar agreements since 2024, renegotiated to align with the new directives from Commerce Secretary Howard Lutnick and the American AI Action Plan.
Trump's Reversal
The Trump administration's narrative on AI was clear until a few months ago: less regulation, more competitiveness. The administration's first act was revoking Biden's executive order on AI safety, dismantling testing and transparency requirements for the most powerful models.
The turning point comes from two converging reasons:
Anthropic's Mythos Model
The announcement of Mythos — Anthropic's model, described as capable of autonomously identifying and exploiting software vulnerabilities — changed the conversation in Washington. An AI system capable of autonomously finding zero-days isn't a consumer product: it's a national security risk. The administration took notice.
Bipartisan Pressure on Cybersecurity
Concerns about malicious use of frontier models — by hostile state actors or criminal groups — found cross-party consensus in Congress. Political pressure pushed the executive branch to act before a real crisis forced a reaction.
What the Agreements Cover
The agreements signed with CAISI aren't law yet. They cover:
- Early access to models before public launch
- Technical evaluations of offensive capabilities, cybersecurity, and dual-use risks
- Collaborative research to develop shared safety standards
Kevin Hassett, director of the White House National Economic Council, stated that the administration is studying an executive order to formalize the process — a "clear road map" for how advanced AI systems should be evaluated before release.
Voluntary Today, Mandatory Tomorrow?
The critical point is the current nature of the agreements: voluntary. Companies agree because they have an interest in maintaining good relations with the administration and avoiding heavier regulation imposed from outside.
The model echoes that of the UK AI Safety Institute — which has built similar agreements with the leading AI labs — and that of Basel for banks: shared technical standards that gradually become industry norm.
If the executive order arrives, participation could become a condition for operating in the federal market — enormous leverage in a sector where government contracts are worth billions.
The Paradox of AI Deregulation
The Trump administration finds itself in a paradoxical position: it dismantled Biden's rules in the name of market freedom, but now finds itself building an oversight system that structurally resembles them — with the difference that it's driven by national security, not consumer protection.
The distinction isn't trivial. Oversight oriented toward national security has different goals than oversight oriented toward transparency or user rights. It filters geopolitical and military risks, not necessarily social or ethical ones.
Conclusion
The US government is formally entering the frontier AI ecosystem, not as an external regulator but as an internal evaluator. Microsoft, Google, and xAI are opening their models before launch — a step that would have seemed impossible twelve months ago. The reason is simple: the power of the most advanced AI models has reached a threshold where even those who believe in deregulation become convinced that someone needs to check what gets released.
Sources: CNN Business · Washington Post · CNBC · Fortune