The head of U.S. Army Cyber Command has disclosed a small task force that builds AI agents for cyber work roles, including agents that now “hunt” in the Department of Defense Information Network (DODIN). Lt. Gen. Christopher Eubank described the effort at the TechNet Augusta conference, DefenseScoop reported.
Task Force Lexington was set up in April to create agents for roles such as developers, data engineers, host analysts and exploitation analysts. “You name the work role, we’re creating,” Eubank said. Agents have already taken on red team roles, and the command now runs 17 agentic mission elements and cyber protection mission elements that scour the DODIN every day.
Small Team, Fast Results
The task force has about 8 to 10 civilians and service members and is led by a lieutenant colonel. Members visited the Defense Innovation Unit in Silicon Valley and came away convinced that the best way to bring AI into the command was to form a small group whose only job is that effort. Eubank said the team is “way ahead” of where he expected, having built and trained agents that mimic human workflows in about 45 days. Its work ranges from helping write situation reports to agentic red teaming, and other units in the command can bring it projects.
Agents are trained “to the same standard that humans get trained,” Eubank said, and their missions are overseen by people. “When they make a mistake, they get taught the way to do it and it goes back in and re-trains itself,” he said.
Guardrails Set Daily
Because agents work far faster than humans, Eubank called the arrangement a “delicate dance.” “Each and every day, we sit down as a group, we figure out the guardrails we’re going to apply to these agents, [and] we ask ourselves: Is it risk that a human should be answering, or is it risk that an agent can answer?” he said. “And right now, today, humans are all responsible for risk. We have not turned any agents loose to assume risk on their own behalf.”
The command has deliberately avoided commercial frontier models, Eubank said, citing token costs and a lack of AI governance. Without better controls, “we’re going to price ourselves out of business,” he said, and he acknowledged that the Army is “never going to have” enough compute.
“We’re now at a place in the cyberspace domain where we have to augment our workforce, and right now the fastest, easiest, smartest, best way to augment the workforce is to create agents and train those agents,” Eubank told DefenseScoop.
For security teams outside the military, the model is a useful template. Narrowly scoped agents trained against defined job roles, corrected and retrained like new staff, and kept on a short leash for risk decisions offer a practical way to add capacity to a stretched security operations team.
WebProNews is an iEntry Publication