Army Cyber Puts AI Agents on Network Duty but Keeps Risk Decisions Human

Army Cyber Command is assigning AI agents to hunt for network threats and intrusions, while reserving risk decisions for human personnel. The approach puts agents into operational cyber roles without authorizing them to act independently when a mission carries risk.

The work is organized through Task Force Lexington, a clearinghouse for Army Cyber Command’s AI efforts. The task force was established in April and has roughly eight to 10 civilian and military members led by a lieutenant colonel, according to Lt. Gen. Christopher Eubank.

That small team is meant to be a common entry point for the command’s AI projects: personnel can bring it a proposed use, and the task force develops the work while tracking its progress. Eubank said the command formed the group after members visited people working on AI and concluded that a focused unit should own the effort.

Defined jobs, not open-ended authority

The command is training agents as developers, data engineers, host analysts and exploitation analysts. It has also assigned agents to network hunting and to support cybersecurity service-provider risk management and enterprise mission-assurance functions.

That framing separates the Army’s program from a model of autonomous cyber action. Agents have work roles, missions and human counterparts; people review the work, then can send an analytic-support agent back on task. The command says agents that make mistakes are taught the correct method and retrained.

Work already assigned to agents

• An agentic red team is in place.
• Agents support cybersecurity service-provider risk management and enterprise mission-assurance functions.
• Agents receive missions under human oversight rather than taking responsibility for mission risk.

17Network-focused mission elements

Army Cyber Command says 17 agentic mission elements and cyber-protection mission elements scan the Department of Defense Information Network every day.

Human qualification is the control point

Army Cyber says its agents are trained to the same standard as people and receive Job Qualification Readiness approval before mission assignment. Human counterparts review their output, send them back on task and retrain them after errors. Eubank said the command has not authorized agents to assume risk independently.

The distinction is consequential for a cyber force trying to move faster without transferring accountability. Eubank said the command sets agent guardrails every day and weighs whether a particular risk belongs with a person or an agent. For now, the answer on responsibility is unambiguous: humans retain it.

Eubank said the task force created and trained agents that copied human workflows in about 45 days. He described setting agent guardrails as a daily exercise because agents move faster than people; the command’s deputy chief of staff for G-6, Lt. Gen. Jeth Rey, said Army defenders must close a machine-speed gap with adversaries.

A constrained model strategy

The command’s push for agent speed is paired with a deliberate choice not to use commercial frontier models. Eubank cited token costs and concerns about AI governance, while saying the work is still supported by industry. His stated concern was that absent governance, AI costs could price the command out of the work.

That leaves Task Force Lexington pursuing a different balance from simply buying access to the largest available models: specialized agents, close human supervision and an internal team building expertise. Eubank also said the command will never have enough compute. The task force has visited the Defense Innovation Unit to meet senior engineers and improve its technical expertise.