Google, Microsoft, xAI Hand US Government Pre-Release Keys to Their AI Models

Google, Microsoft, and xAI have signed voluntary agreements to provide the U.S. government with pre-release access to their frontier artificial-intelligence models, according to multiple outlets citing people familiar with the arrangement. The companies will share model weights, system cards, and red-team evaluation results with the newly created Center for AI Standards and Innovation, a federal body housed within the Department of Commerce. Access begins weeks before any public launch, giving Washington time to probe safety, bias, and misuse vectors. The scope is limited to models deemed dual-use or above a yet-unpublished compute threshold, likely covering GPT-class successors, Gemini Ultra-scale systems, and Grok’s next generation.

Google, Microsoft, and xAI were already the furthest along in voluntary commitments made at last year’s AI Safety Summit, so signing on to an expanded federal preview program was the path of least resistance. Microsoft’s deep Azure cloud contracts with the Pentagon gave it extra incentive to stay ahead of any mandatory regulation. xAI, still private and heavily reliant on Elon Musk’s government relationships via SpaceX and Tesla, saw an opportunity to curry favor. Google, facing antitrust heat on multiple fronts, viewed participation as reputational insurance. Smaller labs like Anthropic and OpenAI reportedly received the same invitation but are still negotiating terms, leaving the initial trio to claim first-mover optics.

The agreement sets a de-facto compliance template that venture-backed startups may soon be asked to follow. Any model trained with more than 10^26 FLOPS could trigger the same pre-release review if federal agencies decide to extend the program. Founders should budget an extra 3-6 weeks for evaluation cycles, build audit trails for training data, and prepare internal red-team documentation. Cloud providers like AWS and CoreWeave will likely add optional “government preview” pipelines to attract labs that want to signal responsibility. Meanwhile, open-source releases could bifurcate: sanitized versions for public GitHub repos and heavier, restricted checkpoints for vetted researchers.

Pentagon officials tell Reuters the program’s primary goal is to catch dangerous emergent capabilities before adversaries weaponize them. The fear is that a single breakthrough in autonomous hacking or bioweapon design could cascade faster than traditional export controls can react. By getting early visibility, the government hopes to craft targeted guardrails instead of sweeping bans. Critics counter that giving Washington a months-long lead on proprietary weights risks leaks or preferential treatment for U.S. defense contractors. The companies insist the data is shared under strict NDAs and that no model details leave the secure evaluation environment.

The U.S. move mirrors the EU’s AI Act pre-deployment conformity assessments but operates on a voluntary basis, avoiding the legislative gridlock that killed earlier Senate bills. Britain’s AI Safety Institute already runs similar private evaluations, and Canada is drafting comparable rules. China’s draft measures require domestic labs to file security assessments with regulators, though with far less transparency. Washington’s approach keeps the innovation edge while signaling to allies that the U.S. can police its own champions without export bans. Expect G7 discussions next month to push for mutual recognition of these early-access protocols, effectively creating a Western AI inspection regime.

OpenAI and Anthropic are expected to join the program within weeks, according to Bloomberg. The Commerce Department will publish a formal framework by July that defines compute thresholds, disclosure timelines, and penalties for non-compliance. Congressional staffers tell The Verge that voluntary participation could become a prerequisite for federal cloud credits, turning an opt-in handshake into a soft mandate. Venture firms are already inserting “reg-ready” clauses into term sheets, and insurers like Munich Re are drafting AI-specific liability riders tied to government review status. The first public test case arrives this fall when Google ships Gemini 2.0; if the model passes federal scrutiny without leaks or delays, the template becomes the new normal.