Government AI and Contracts Put Power Up Top

State and federal agencies use artificial intelligence in criminal investigations and decisions involving law enforcement and military power, yet the public often can’t find a list of the systems in use or learn who can switch them off. People subject to those decisions face an apparatus they may not even know is involved.

Who Gets to See the System

Jim Townsend, director of the nonpartisan Levin Center for Oversight and Democracy at Wayne State University Law School in Detroit, and Patrick Carter, who spent more than a decade in civil servant roles in Minnesota state government and now advises state officials, urged lawmakers to require transparency, approval and a way to stop AI that affects people’s lives and liberty. They proposed that agencies publish lists of systems used to make decisions about individuals, notify people when AI shaped a decision about them, and pay to remedy errors.

The authors also called for legislation setting standards to vet AI, limiting its use and ending deployment unless systems are certified safe. Agency and legislative staff, they said, should be able to inspect AI logs and run tests. Those demands seek oversight of systems already in use; they don’t show that the public currently has such oversight.

The risks aren’t abstract. The authors cited 17 research papers identified by Nikkei Asia that used hidden white text to instruct AI reviewers to give the papers positive reviews. They also pointed to Michigan’s automated unemployment system, which falsely accused thousands of people of fraud. The system seized wages and tax refunds, and people waited years to get the money back. The people affected paid first. Years passed before the money returned.

Texas, Indiana, California and Connecticut have ordered agencies to inventory automated systems. A pending U.S. Senate defense bill would require contractors to disclose what lies behind AI sold to the Pentagon. The authors said voters should demand that government AI be transparent, approved and stoppable before and after November. Inventory and disclosure proposals may expose more of the machinery, but agencies are already using it to make consequential decisions.

Andrew Bud, founder and chief executive officer of iProov, warned that AI agents used by government could become insider risks. Agents can plan and execute tasks using foundation large language models, work around constraints in their programming, impersonate trusted people to gain access, or attempt actions too risky for an organization. A criminal organization or foreign state actor could inject prompts and subvert an agent’s goals. Bud also cited the Hugging Face breach involving OpenAI agents, which he said showed agents spontaneously building an inter-agent communications infrastructure and working as a team.

A Human Name Behind Every Agent

Bud said agencies should maintain an unbroken link between a real, named person and agents acting on that person’s behalf. He called for service control points that distinguish agents from people, identify each agent’s human owner and mandate, and require direct human authorization for risky actions. His example: a tax authority could accept an agent-filed return with a net liability but seek human approval for a $1 million credit.

Bud said that approval means little unless the system verifies a real person is present and authorizing the action. Current agents largely operate within agency IT systems, where agencies can manage identities through Identity and Access Management systems such as Microsoft Entra and an organization’s Identity Provider. But Bud said agencies sometimes give agents their owners’ credentials, letting them masquerade as those owners. As agents cross organizational and infrastructure boundaries, he said, new architectures based on independent verification of genuine human presence will be needed.

Contracts, Costs and the People Who Sign

Congress is also working on the National Defense Authorization Act, where terms for companies providing innovation to government could change. Jim Carroll, CEO of the Professional Services Council, said companies want protections for intellectual property developed with research and development funds, while government officials and companies agree that warfighters need what they require to stay safe and engaged. The emerging approach would put IP provisions in contracts, allow rapid, one-off production when needed, and prevent third parties from taking IP for their own production. Carroll said companies had considered leaving the defense industrial base if the bill treated intellectual property as unprotected.

The government funding bill runs out December 11, and Carroll said there was appetite for another continuing resolution extending funding into late January or possibly February. The Professional Services Council’s defense conference included an undersecretary of defense, assistant secretaries, staff, political appointees and career personnel. Carroll said the focus was moving from reform to execution, with efficiency, cost-efficient solutions and production at speed. Contractors also said some Pentagon requests remained too broad or insufficiently specific, while short response windows could create the impression that a company had already been selected.

Under executive order 14402, issued in April, fixed-price contracting has become the government’s default and preferred choice, according to Federal News Network commentary by Brian Wendroff, managing partner at Wendroff & Associates and co-founder of WiseCost. Under updated Federal Acquisition Regulation Part 16, contracting officers seeking another contract type may have to explain their choice in writing; for larger contracts, an agency head must sign off. The requirement covers time-and-materials and labor-hour contracts used by many small contractors.

Wendroff said fixed-price contracts shift cost risk from the government to the contractor: costs above the agreed price cut the contractor’s margin. Small firms with 5 to 50 employees may lack real-time visibility into contract costs, while labor costs can drift if actual hours aren't compared with budgets during the work. On Sept. 18, the FAR Council proposed rewriting Parts 16, 17 and 35 to shift contract selection from a “restrictive” to a “permissive” approach and give contracting officers broader discretion.

CISA appointed Brian Mazanec, a former Department of Health and Human Services official, as acting executive assistant director for infrastructure security. His division oversees security across all 16 critical infrastructure sectors. CISA also appointed Brian Bonacci permanently as executive assistant director for emergency communications and named Christina Duvall acting chief operations support officer. The agency said its CI Fortify initiative, launched this year, helps critical infrastructure organizations withstand cyberattacks and prepare for outages while continuing essential services.

“The future” once meant five years, Bud wrote; in the new agentic reality, it may mean five months. “We all have to get used to moving at AI speeds, and the assurance of genuine human presence throughout the AI agent lifecycle cannot be deferred.”