The news: Google DeepMind, Microsoft, and xAI signed agreements with the Center for AI Standards and Innovation (CAISI) to let the US government evaluate their AI models before public release. CAISI has already completed 40+ model evaluations.

The context: This comes 24 hours after reports that the White House is weighing an executive order on AI oversight. Anthropic's Mythos model — explicitly designed for cyber operations — reportedly tipped the scales. The government wants to see frontier models before they're deployed.

What's in the agreement:

  • Pre-deployment access to unreleased models
  • Post-deployment monitoring and testing
  • CAISI evaluates national security and public safety implications
  • "Technical, scientific and national security expertise" from the Commerce Department

Who's missing: OpenAI and Anthropic. OpenAI already said last week it's giving "all vetted levels of government" access to its most advanced models. Anthropic is restricting Mythos to approved organizations and briefing senior officials directly. Both are already cooperating — just not through this specific framework.

Why this matters:

  1. Industry standardization: If three of the biggest labs sign up, the rest will face pressure to follow. CAISI becomes the de facto gatekeeper for frontier model releases.

  2. The Anthropic problem: The government wants Anthropic's cyber capabilities but excluded Anthropic from classified Pentagon deals. CAISI offers a workaround — oversight without partnership.

  3. Global precedent: If the US claims pre-release review rights, the EU, UK, and China will demand the same. This could become the template for global AI governance.

The tension: The Trump administration has spent months rolling back AI safety regulations. Now it's building a review process after a single model scared enough people. This isn't principled regulation — it's reactive risk management.

Bottom line: CAISI just became the most powerful AI regulator in the world, and it didn't exist six months ago. The labs are volunteering for oversight because the alternative — executive orders, congressional action, or export controls — looks worse.