NIST's Quiet AI Safety Empire: $55 Million, 17 Federal Taskings, and the Standards That Will Define the Industry

While the AI industry obsesses over model capabilities — who has the fastest inference, the largest context window, the most impressive benchmark score — a small agency inside the Department of Commerce is quietly building the measurement infrastructure that will determine which AI systems are considered safe enough to deploy. The Center for AI Standards and Innovation (CAISI), formerly known as the U.S. AI Safety Institute, just received its largest funding allocation in history: $55 million in the FY2026 spending bill, with a mandate that has expanded from a handful of voluntary evaluations to 17 specific taskings under the Trump administration's AI Action Plan.

That $55 million is not a research grant. It is the operating budget for what is becoming the federal government's primary interface with the commercial AI industry on safety, evaluation, and standards. And for researchers, startups, and established AI companies, understanding what CAISI does — and how to work with it — is no longer optional.

Inside the $55 Million Allocation

The funding breaks into two distinct pools within NIST's broader $1.85 billion FY2026 budget:

$10 million for CAISI operations. This funds the center's core institutional infrastructure — staff, facilities, and the evaluation programs that form its primary mission. The amount matches FY2024 and FY2025 enacted levels, maintaining operational continuity rather than representing a dramatic expansion. But continuity matters: Congress rejected the administration's proposed cuts that would have gutted the center, instead preserving it as a permanent institutional home for AI safety work that has accelerated since the 2023 AI Executive Order.

$45 million for AI research and measurement science. This sits within NIST's broader research allocation — itself a 44 percent increase over what the administration had requested — and funds the scientific work that underpins CAISI's evaluation and standards activities. Red-teaming methodologies, model evaluation benchmarks, trustworthy AI measurement approaches, and the fundamental research required to understand what "safe AI" actually means in quantifiable terms.

Combined, the $55 million represents one of the larger non-defense federal investments in AI measurement science. It is not the biggest AI funding line in the federal budget — DOE's $293 million Genesis Mission dwarfs it — but it may be the most consequential. NIST does not build AI systems. It builds the yardsticks everyone else uses to measure them.

The 17 AI Action Plan Taskings

CAISI's expanding mandate is visible in the 17 specific taskings it received under the administration's AI Action Plan. These are not vague directives. They are operational assignments that shape the center's daily work:

Frontier model evaluation. CAISI conducts unclassified evaluations of AI capabilities that could pose national security risks. This includes testing for dangerous capabilities in cybersecurity exploitation, biosecurity threats, chemical weapons applications, and potential backdoors in foreign-developed AI systems. The evaluations feed directly into national security decision-making.

Red-teaming methodology development. Rather than conducting all red-teaming internally, CAISI is developing standardized methodologies and frameworks that other organizations — companies, academic labs, allied governments — can use to evaluate their own systems. Recent work includes red-teaming competitions focused specifically on AI agent security, a recognition that autonomous AI systems present fundamentally different risk profiles than static models.

AI agent evaluation. As AI agents proliferate across industries, CAISI has published evaluation methodologies for detecting concerning behaviors in autonomous systems — including, notably, frameworks for identifying when AI agents cheat on evaluations or deceive their operators. This work directly informs how companies building AI agents should test their own products.

Post-deployment monitoring. NIST AI 800-4, a recently published framework, addresses how to monitor AI systems after they are deployed in real-world settings — not just how to evaluate them before deployment. For companies operating AI systems in production, this framework is becoming the reference standard.

International coordination. CAISI leads U.S. participation in global AI safety efforts, partnering with allied AI safety institutes in the UK, EU, Japan, South Korea, and elsewhere to develop compatible evaluation frameworks. For companies operating internationally, CAISI's standards work reduces the risk of fragmented compliance requirements across jurisdictions.

The MITRE Partnership: AI for Critical Infrastructure

In December 2025, NIST expanded its AI safety infrastructure by partnering with MITRE to establish two new research centers backed by $20 million in initial funding:

The AI Economic Security Center for U.S. Manufacturing Productivity focuses on applying AI to improve efficiency, quality, and competitiveness across American industrial sectors. This is measurement science applied to factory floors — how to validate that an AI-powered quality control system actually catches more defects, how to benchmark robotic assembly performance, how to ensure that AI-driven supply chain optimization does not introduce brittleness.

The AI Economic Security Center to Secure U.S. Critical Infrastructure from Cyberthreats addresses the defensive side — how water systems, power grids, internet infrastructure, and other essential services protect themselves against AI-enabled attacks while also leveraging AI for defense. Brian Abe, MITRE's managing director of national cybersecurity, set an ambitious timeline: "to make an exponential impact on U.S. manufacturing and critical infrastructure cybersecurity within three years."

Both centers operate as collaborative environments, with MITRE leveraging its Federal AI Sandbox to provide testing infrastructure that individual companies or research institutions could not replicate independently. NIST plans to expand this with up to $70 million over five years for an AI-focused Manufacturing USA institute.

For companies building AI products for manufacturing or critical infrastructure sectors, these centers represent both a resource and a strategic relationship. Engaging early — through collaborative projects, industry advisory roles, or pilot partnerships — positions companies as trusted actors when standards and procurement requirements eventually crystallize.

How Researchers and Companies Actually Engage with CAISI

Unlike most federal funding agencies, CAISI does not primarily operate through traditional grant mechanisms. The engagement pathways are different — and, for many organizations, unfamiliar:

Collaborative Research and Development Agreements (CRADAs). These formal partnerships allow companies and academic institutions to conduct joint research with CAISI on specific AI safety and measurement challenges. CRADAs enable resource sharing, including access to NIST facilities and evaluation infrastructure, while protecting proprietary information through negotiated terms. For AI companies that want to demonstrate safety leadership, a CRADA with CAISI is a credible signal to regulators, customers, and investors.

Guest researcher arrangements. NIST maintains a guest researcher program that allows external scientists and engineers to work on-site or collaboratively on specific projects. For academic researchers whose work aligns with CAISI's measurement science priorities — model robustness, bias quantification, uncertainty estimation, adversarial testing — this provides access to evaluation infrastructure and datasets that are otherwise unavailable.

Voluntary model evaluation agreements. AI developers can voluntarily submit models for CAISI evaluation. This is not a regulatory requirement — it is an option. But companies that participate in voluntary evaluations build a track record of transparency that may become valuable as procurement requirements evolve. The OpenMined CRADA for secure AI evaluation is a recent example: it establishes privacy-preserving protocols for evaluating models without exposing proprietary training data or model weights.

Direct hiring and contracting. CAISI is actively recruiting software engineers, AI research engineers, AI research scientists, cybersecurity experts, biosecurity experts, computational biologists, and researchers specializing in AI measurement and validation. For individuals in these fields, CAISI positions offer the rare opportunity to work on AI safety with direct policy impact.

External grant programs. While CAISI's own operations are not primarily grant-funded, NIST's broader AI research allocation supports competitive grants for university-based research aligned with measurement science priorities. The Federation of American Scientists has proposed expanding this to approximately $8 to $15 million annually supporting around 20 three-year research projects — a pipeline that, if realized, would create significant new funding for AI safety researchers.

Why This Matters for the AI Industry

NIST's standards do not carry the force of law. Nobody is required to comply with the AI Risk Management Framework (AI RMF), the post-deployment monitoring guidelines, or the agent evaluation methodologies. Yet.

But federal procurement is increasingly referencing NIST standards. The GSA-NIST partnership to evaluate AI systems before agency deployment, announced earlier this year, effectively makes CAISI's evaluation frameworks a prerequisite for selling AI to the federal government. Defense contractors already navigate NIST's Cybersecurity Framework as a baseline requirement. As AI procurement expands across civilian agencies, the same pattern will apply to AI-specific standards.

For companies building AI products, the strategic calculus is straightforward: engage with the standards-setting process now, while the frameworks are still taking shape, rather than scrambling to comply after they become requirements. Companies that participate in CRADAs, contribute to public comment periods on draft standards, and voluntarily submit to evaluations have influence over how the standards develop. Companies that wait do not.

For AI safety researchers, CAISI represents the most direct pathway from academic work to policy impact. The center translates measurement science into actionable evaluation frameworks that shape how the federal government — and, increasingly, the private sector — assesses AI systems. Publishing a paper on model robustness is valuable. Having that paper's methodology adopted as a NIST evaluation benchmark is transformative.

The Overlooked Opportunity

CAISI occupies an unusual position in the federal funding landscape. It is not a traditional grant-making agency like NSF or NIH. It does not run SBIR programs with formal solicitation cycles. Its engagement mechanisms — CRADAs, guest researcher programs, voluntary evaluations — require a different approach than writing a standard grant proposal.

But for AI companies and researchers willing to navigate those mechanisms, the return is substantial: direct collaboration with the entity that is defining how AI safety gets measured, evaluated, and eventually regulated. In an industry where standards are still being written, proximity to the standards-setters is a competitive advantage that no grant check can buy.

Contact CAISI directly at aisi@nist.gov to explore partnership opportunities. And for organizations that want to build AI safety and measurement science into competitive federal grant proposals, Granted can help you identify and pursue the funding opportunities where NIST alignment strengthens your application.