White House blocks CAISI testing deals with Google and Microsoft

The intervention signals a centralized shift in how the U.S. government intends to audit frontier models, moving away from independent agency agreements toward a unified executive framework. For the streaming and tech ecosystem, this regulatory friction complicates the deployment of advanced AI agents used in content recommendation and automated production, as developers face inconsistent safety standards across different federal departments. The lack of a permanent CAISI director and the ongoing 'turf war' between the Commerce Department and the NSA may delay the establishment of clear compliance benchmarks. Watch for the introduction of major bipartisan AI safety legislation in Congress within the next eight months to provide a statutory baseline for these testing protocols.

The Center for AI Standards and Innovation (CAISI) has operated without a permanent director since its establishment under the Commerce Department, creating a leadership vacuum that complicates its ability to finalize voluntary testing agreements with frontier AI developers. Google and Microsoft have both maintained public commitments to third-party safety evaluation, with Google publishing its Frontier Safety Framework in May 2025 that outlines critical capability levels and mitigation strategies for models approaching dangerous thresholds. Microsoft similarly released its Responsible AI Standard and has participated in multiple government-led evaluation initiatives, though the company has not publicly commented on the specific CAISI agreements blocked by the White House intervention.

The broader regulatory landscape for AI testing is fragmenting across multiple federal agencies and legislative proposals. The National Institute of Standards and Technology released its AI Risk Management Framework in January 2023, which has become the de facto reference for voluntary AI governance across industry, but it lacks enforcement mechanisms. Meanwhile, the EU AI Act entered into force in August 2024 with phased compliance deadlines, creating a divergent international standard that U.S. developers like OpenAI, Anthropic, and Meta must navigate alongside domestic uncertainty. The EU AI Act's general-purpose AI obligations began applying in August 2025, requiring transparency documentation and systemic risk assessments for models trained with more than 10^25 FLOPs, a threshold that captures most frontier systems from the companies named in the CAISI dispute.

On the technical side, independent AI evaluation organizations have stepped into the gap left by stalled government testing programs. METR published findings in March 2025 showing that frontier models can autonomously complete software engineering tasks lasting up to 50 minutes, a benchmark that has become a reference point for assessing model capability growth rates relevant to safety thresholds. The organization's time-horizon metric, which measures how long a model can sustain autonomous task completion, has been cited by multiple government bodies as a potential basis for standardized testing protocols. For streaming companies deploying AI-driven content recommendation and automated production tools, the absence of a unified U.S. testing standard means compliance obligations may shift rapidly depending on which agency or executive action ultimately prevails.