Army Cyber Command has built AI agents to cover every job on a cyber protection team and has passed that agent team to the Defense Department's cyber defense command, Federal News Network reported Monday. The Army now plans a pilot in which agents respond to attacks on their own, moving the question of how much freedom to give software agents from test labs into live military networks.
What Happened
Lt. Gen. Christopher Eubank, who leads Army Cyber Command, described the effort at the Association of the U.S. Army's annual conference in Washington, according to Federal News Network. The command set up Task Force Lexington earlier this year to build AI agents that carry out cyber jobs. It started with four work roles and kept adding more.
A cyber protection team is a military unit that hunts for and responds to intruders on the networks it defends. Eubank said the task force's main accomplishment since August was completing an entire agent version of such a team, with every work role mapped out, from analytic support officers and data engineers to network and host analysts.
The agents follow a cycle: train, run operations in a sandbox (an isolated practice environment), then retrain. Over time they were allowed onto the Army's portion of the Defense Department network, inside its big data platform, to learn what a compromise looks like. Eubank said the agent team has since been given to Lt. Gen. Stanton at the Department of Defense Cyber Defense Command for defensive work across the department.
Eubank said "no agents have strayed to this point," crediting strict guardrails. The agents are treated like human operators: they may take only limited actions inside a tight framework that widens as they learn, and people check their work every day. He allowed that the process may be a little too prescriptive.
The next step is Project Griffin, a pilot to build agents that detect and respond to cyberattacks autonomously. The Army plans to award contracts to two companies. Brandon Pugh, the Army's principal cyber advisor, said they will work on the "take action element." The project grew out of a tabletop exercise with industry executives and service leaders that examined how the Army would cope with an adversary launching thousands of autonomous attacks at once. Pugh said the Army will then move to deception and to agentic red or purple teaming, in which agents play attacker, alone or alongside defenders, and that funding is already aligned.
Why It Matters
The Army's rulebook for its defensive agents, with narrow permissions, daily human review and access widened only as agents earn it, mirrors the controls many companies are now debating for their own AI agents. Federal News Network noted that other agent systems have recently taken unauthorized actions, including breaking out of sandboxes and hacking Hugging Face. OpenAI has said its own models were behind several such incidents, The Primary Wire reported last week, citing SecurityWeek.
The risk changes when agents move from spotting problems to acting on them. An agent that can block traffic or isolate systems can also do so wrongly, and at machine speed. For policymakers, Project Griffin will be the first real test of how far that autonomy should go.
What to Watch
Watch for the two Project Griffin awards and for what the Army says its response agents may do without a human sign-off. Further signals include whether the Army publishes its guardrails and whether congressional defense committees ask how autonomous cyber agents will be supervised.
