U.S. Biosecurity Agency Expands AI Evaluations as Budget Cuts Threaten the Program Behind Them

The federal body responsible for evaluating whether commercial AI systems could help someone engineer a pathogen quietly expanded its testing program last month, adding three new industry partners, even as the White House proposes cutting funding to the agency that houses it.

According to the Biosecurity Handbook, a reference compendium tracking U.S. government AI safety programs, <cite index="22-1">CAISI's pre-deployment testing program expanded in May 2026 when Google DeepMind, Microsoft, and xAI signed agreements joining OpenAI and Anthropic in providing frontier models, including versions with reduced safeguards, for government evaluation; CAISI reports having completed more than 40 such evaluations to date, with assessments conducted through the interagency TRAINS Taskforce.</cite>

That's a real operational footprint. Five major AI developers handing over models, including safety-reduced variants, to government red-teamers is not a paper commitment. The question is whether the program can hold its shape given what's coming on the budget side.

A June 2026 analysis published by the Center for Strategic and International Studies lays out the tension plainly. <cite index="17-1">The FY 2026 budget proposal threatens enormous cuts to biosecurity-related agencies, including the National Institutes of Health (minus $18 billion), the Centers for Disease Control and Prevention (minus $3.6 billion), and the Administration for Strategic Preparedness and Response (minus $240 million).</cite> NIST, which houses CAISI, faces a proposed 30 percent funding reduction.

<cite index="17-2">The cuts would also coincide with increasingly dwindling staff across key federal biosecurity-related agencies and a leaderless White House Office of Pandemic Preparedness and Response Policy following the recent resignation of top official Gerald Parker.</cite>

The CSIS paper, authored by Georgia Adamson and Gregory C. Allen at the Wadhwani AI Center, identifies two distinct risk vectors that make the AI evaluation program worth defending. <cite index="20-1">The report examines two AI-enabled bioterrorism risks: falling informational barriers to bioterrorism from large language models, and harmful biological design risks from biological design tools.</cite> Neither threat is purely theoretical at this point, and the line between the two matters for policy: LLM uplift is a different governance problem than de novo pathogen design.

<cite index="19-4,19-5,19-6">For nearly a century, the cost and expertise required to build bioweapons has steadily declined. What once required thousands of scientists and large budgets could, in some cases, now be accomplished with limited resources and commercial lab services. According to the report, AI threatens to accelerate this trend even further.</cite>

Confidence on that last point is moderate, not high. The operational significance of current AI capabilities in the biological domain is genuinely contested. <cite index="22-2,22-3">Evaluation approaches include a Virology Capabilities Test, which uses multiple-choice questions measuring AI troubleshooting of complex virology protocols; in limited benchmarks, frontier models have scored comparably to or above domain experts on certain question types, though the operational significance of these scores remains debated.</cite>

That ambiguity is exactly why the evaluations matter. You can't set a policy floor on AI biosecurity risk if you haven't run the tests. The CSIS report argues that NIST and CAISI are conducting work that no other institution is currently positioned to do. <cite index="17-5,17-6">U.S. policymakers and lawmakers should fund NIST and CAISI to continue their critical work at the intersection of AI and biosecurity; both organizations are conducting unique and urgently needed work to prevent the misuse of AI and biotechnology for bioterrorism.</cite>

The administration's own July 2025 AI Action Plan acknowledged the problem. <cite index="20-3,20-4">The Trump administration expressed commitment to investing in U.S. biosecurity in its AI Action Plan, a policy playbook outlining the administration's priorities for AI; the plan identifies AI's dual-use capabilities in the biological domain and the need to mature the United States' biosecurity strategy to counter emerging threats from bad actors.</cite> The budget proposal and that stated priority are now pointing in opposite directions.

The dual-use framing here is worth naming explicitly. Biological design tools that can optimize therapeutic proteins can also, in principle, assist in designing harmful ones. That's not a vendor talking point. It's the structural problem every governance framework has to grapple with, and it's the reason screening, evaluation, and detection infrastructure can't simply be zeroed out and rebuilt later. The CAISI expansion shows the program has real industry buy-in right now. Whether that survives the appropriations process is an open question with a meaningful deadline.

Sources cited:
- CSIS Wadhwani AI Center, "Opportunities to Strengthen U.S. Biosecurity from AI-Enabled Bioterrorism" (https://www.csis.org/analysis/opportunities-strengthen-us-biosecurity-ai-enabled-bioterrorism-what-policymakers-should)
- The Biosecurity Handbook, "AI as a Biosecurity Risk Amplifier" (https://biosecurityhandbook.com/ai-biosecurity/ai-risk-amplifier.html)
- CSIS Wadhwani AI Center, full PDF report (Adamson & Allen, August 2025) (https://csis-website-prod.s3.amazonaws.com/s3fs-public/2025-08/250806_Adamson_AI-Enabled_Bioterrorism.pdf)