USDA cannot keep a person in the loop for all AI tasks

Rudolf Rojas is the security lead at the US Department of Agriculture (USDA), in Washington. Rojas told a summit that cuts removed the resources and contract support to keep a person in the loop for AI tasks. Rojas wants agents to do the work and report back, but gave no test results, and other officials did not agree.

How it works

Rojas wants agents with “minimal privileges, minimal role”. Agents can work without a person only with a risk plan and a check of the networks. They also need a demilitarized zone (a network area between the inner and outer networks). Zero-trust rules (the system checks each access) let USDA monitor networks and tools.

The numbers

Rojas gave no safety numbers. Two officials gave numbers on use:

• Gray Barrett of the State Department estimates that its chatbot saves two hours a week.
• Bill Watkins of the National Nuclear Security Administration reports waits of six months for some hardware.

What is not known

Ajay Phogat of Customs and Border Protection is for a person in the loop. Kevin Walsh of the Government Accountability Office told that the government cannot keep up and has not sufficient oversight. Rojas wants national rules and does not fully agree with the administration on “free rein” for AI firms.

Insight: The person moves from each task to the rules

Rojas removes the person who checks each task. The Atlassian CEO tells that each agent must have a person as owner. Two AWS engineers use code hooks to stop a defective agent. This reads as a move of the person from task checks to owner of the agent and its limits. Next, look at who writes those limits: Rojas tells that vendors must make their tools better.