Every Major US AI Model Now Goes Through Government Vetting
In a landmark shift for AI governance in the United States, the Commerce Department's Center for AI Safety and Innovation (CAISI) has finalized pre-deployment evaluation agreements with all five major frontier AI labs: OpenAI, Anthropic, Google DeepMind, Microsoft, and xAI. The agreements, finalized in May 2026, mean that every major new AI model released in the United States must now undergo a government safety evaluation before public launch — a requirement with no equivalent in AI's previous history.
The framework stops well short of a formal regulatory approval process: CAISI evaluations are not "approvals" in the pharmaceutical sense, and labs are not legally prohibited from launching models that receive a concerning evaluation. However, the agreements create a powerful accountability mechanism. Labs that launch models despite CAISI concerns would be doing so in the face of a documented government safety finding — a reputational and potentially legal risk that most companies are unlikely to take.
What CAISI Actually Tests
The CAISI evaluation framework focuses on four categories of potential harm: biological weapons uplift (whether a model could meaningfully assist a bad actor in synthesizing dangerous pathogens), cyberweapon capability (whether a model could enable novel cyberattacks), critical infrastructure vulnerability (whether a model could be used to attack power grids, financial systems, or transportation networks), and democratic integrity (whether a model creates capabilities for large-scale election manipulation or disinformation campaigns).
Labs submit their models for evaluation approximately 90 days before planned launch. CAISI employs a team of "red teamers" — professional adversarial testers — who attempt to extract dangerous capabilities using both standard prompting and specialized attack techniques. The results are shared with the lab in confidence, and a public summary is published after launch describing the evaluation's scope and general findings, without revealing proprietary model information.
How US Labs Are Responding
All five labs have integrated the CAISI evaluation timeline into their product roadmaps. Anthropic, which has long invested heavily in AI safety research, described the framework as "broadly aligned with our internal evaluation practices." OpenAI called the agreement "an important step toward responsible AI deployment at scale." Google DeepMind and Microsoft emphasized the framework's role in maintaining public trust in AI systems as they become more capable.
Behind the public statements, however, the evaluations have created new operational pressure. The 90-day window means that competitive model releases — where labs have historically tried to launch as quickly as possible to capture market share — now require earlier internal readiness milestones. Labs that fall behind on safety evaluation timelines risk delaying product launches that have significant commercial implications.
International Implications: The Race to Regulate AI
The US framework is being watched closely by regulators in the European Union, the United Kingdom, Japan, and China. The EU AI Act entered full enforcement in August 2026, imposing mandatory compliance requirements on high-risk AI systems across all EU member states. The UK's approach remains more principles-based, with sector regulators taking the lead rather than a single AI-specific authority. China has its own pre-deployment review system for generative AI, implemented through the Cyberspace Administration of China.
The result is an increasingly complex patchwork of AI governance requirements that multinational technology companies must navigate. Labs launching models globally now must satisfy CAISI evaluation timelines in the US, EU AI Act conformity assessments in Europe, and CAC review processes in China — often on products that are architecturally identical but may need to be evaluated separately under each jurisdiction's criteria.
What This Means for Enterprise AI Procurement
For US enterprises procuring AI systems, the CAISI framework provides a new layer of due diligence infrastructure. Models that have successfully completed a CAISI evaluation carry an implicit endorsement of baseline safety — a signal that is likely to become a procurement requirement in regulated sectors including federal contracting, healthcare, and financial services. CISOs and legal teams advising on AI vendor selection should begin asking vendors to document their CAISI evaluation history as part of standard supplier qualification processes.
