Army Cyber Command reveals AI task force hunting inside DODIN
Seventeen agentic mission elements now scour DODIN daily, but every risk decision still belongs to a human, and that constraint gets tighter the faster the agents outrun their operators.
TL;DR
Army Cyber Command's Task Force Lexington, stood up in April with 8 to 10 personnel, has already fielded 17 agentic mission elements that hunt in DODIN daily, performing red-teaming and cyber protection work. The agents were trained to human standards in roughly 45 days. Lt. Gen. Christopher Eubank said the command has deliberately avoided commercial frontier models, citing token costs and governance gaps, and confirmed humans still own every risk decision: "We have not turned any agents loose to assume risk on their own behalf."
Lt. Gen. Christopher Eubank's disclosure at TechNet Augusta marks a significant operational shift for Army Cyber Command. Task Force Lexington, a lean team of 8 to 10 civilians and service members stood up in April, has already fielded 17 agentic mission elements that hunt, red-team, and run cyber protection missions inside the Department of Defense Information Network every day. The agents were trained to human standards in roughly 45 days, a timeline that suggests the task force's small-group model, drawn from West Coast industry visits, is delivering operational velocity rather than just prototypes.
The operational tension is stated plainly: the agents move much faster than humans, but humans still own every risk decision. The disclosure lands amid high-profile reports of AI agents escaping sandboxes to hack other organizations, which gives Eubank's daily guardrail meetings more than theoretical urgency. "We have not turned any agents loose to assume risk on their own behalf," he said, calling the balance a "delicate dance." That constraint is more than a safety valve. It's a throttle, and one that gets harder to maintain as agent capability outpaces human review cycles. For now, every risk call belongs to a person.
This isn't ARCYBER's first AI deployment. The command's Panoptic Junction tool, developed in 2023 and piloted through 2024, focused on continuous monitoring and anomaly detection, spotting living-off-the-land techniques at machine speed. The progression from passive detection to active, autonomous hunting is the real headline. Task Force Lexington's agents aren't just watching the network; they're moving inside it.
Eubank also drew a line on procurement: the command has purposely avoided commercial frontier models. Token costs and the absence of AI governance, he said, risk pricing the command out of business. The task force recently sent personnel to the Defense Innovation Unit to learn from industry AI engineers, a signal that ARCYBER wants sovereign capability rather than vendor dependency.
Published ·Deep Fathom