nist-800-171/trade-press/
USDA OIG finds AI systems deployed without required ATOs or governance
The department has a chief AI officer and an AI inventory. It doesn't have policies, risk controls, or authority-to-operate approvals for almost any of it.
Editorial brief
A USDA OIG report released last week found the Agriculture Department has deployed AI across supply chain risk, crop yield estimation, and permitting without completing required cybersecurity and governance controls. Almost none of the AI use cases in the FY2024 inventory carried an authority to operate. USDA has no generative AI policy, hasn't updated agency-level AI policies, and hasn't implemented minimum risk management practices for high-risk AI systems, as required by OMB guidance. The OIG also flagged shadow AI risk, noting the department relies solely on an annual employee self-report to track AI use. USDA agreed with all recommendations.
The USDA OIG report lands in an uncomfortable place for an administration that has made aggressive AI deployment a stated priority: the department moved fast on the technology and, by the OIG's account, skipped most of the compliance infrastructure that federal standards require.
The specific gaps are procedural and consequential. Authority-to-operate requirements exist precisely to force a documented risk review before a system goes live. The OIG found that almost none of USDA's FY2024 AI inventory had one. Without ATOs, there is no formal assurance that the systems meet FISMA-required cybersecurity controls, and, as the report notes, management has no documented basis to claim otherwise.
The OMB framework at issue blends Biden-era AI governance guidance (OMB M-24-10) as modified under the Trump administration. USDA has installed the required chief AI officer, which satisfies one structural requirement, but hasn't followed through on the policy and risk-management obligations that accompany that role. High-risk AI systems (those touching civil rights determinations or critical infrastructure, both categories explicitly flagged in OMB guidance) are supposed to have minimum risk management practices in place. The OIG found they don't.
The shadow AI finding deserves separate attention. An annual self-reported data call is a weak control for any technology environment, but it is a particularly weak one for AI, where adoption is fast, tooling is embedded in commercial software employees already use, and the line between sanctioned and unsanctioned use is blurry by design. The OIG flagged this explicitly as a gap; the department agreed.
USDA agreed with all recommendations in the report. Agreement is the easy part. The harder question (not resolved by this report) is sequencing: how a department that has already deployed AI in operationally consequential workflows retrofits governance controls without disrupting those workflows. That implementation problem is USDA's to solve, and the OIG will presumably check progress in a follow-on review.
Contractors and integrators supporting USDA AI systems should note the ATO gap specifically. Systems without an ATO are systems without a confirmed control baseline, and that has direct implications for any NIST SP 800-171 or FISMA representations made in support of those contracts.
Published ·Updated ·Deep Fathom