\u201cThe task force has built those out. They\u2019ve been training them, and then having them do operations inside of a sandbox,\u201d Eubank said. He added that humans oversee the agents\u2019 actions, retrain them and then have let the agents work inside the Army\u2019s network \u201cover time.\u201d\u00a0\u00a0<\/p>\n
The Army\u2019s development of AI agents for cyber purposes comes amid deep concerns about the technology, especially after high-profile incidents where agents from frontier labs have escaped human oversight and hacked third-party organizations.\u00a0<\/p>\n
At the same time, service officials have said that adversaries are wielding the technology against the U.S., and it is too powerful to sideline against those threats.\u00a0<\/p>\n
Eubank said none of ARCYBER\u2019s agents \u201chave strayed to this point,\u201d citing the command\u2019s \u201ctight framework\u201d on using them.<\/p>\n
\u201cWe\u2019ve been very deliberate in the guardrails we\u2019re putting in place,\u201d he said. \u201cWe\u2019re treating each one of these agents just like a human operator, so they are basic. And so with that mindset, they\u2019re only allowed to do so much.\u201d<\/p>\n
Eubank described a process in which human operators check agent work in the network and retrain them before the human-machine pair moves on to the next task. At its outset, Task Force Lexington was staffed by nearly a dozen civilians and service members, and headed by a colonel.<\/p>\n
While the roles are narrow, Eubank alluded to even more expanded use of agents and said that the task force is sharing tools it has built to other cyber-oriented military agencies such as the Defense Cyber Defense Command (DCDC).\u00a0<\/p>\n
\u201cAs they get trained and learn, we open the framework up,\u201d he said of the AI agents. \u201cThe guardrails are really tied to cybersecurity, really based on what humans have learned over the years on what we should do inside of operating and securing the network, defending the network, and attacking a network. We have been very deliberate in kind of \u2018okay, what did we learn as humans? How do we write those guardrails, and then how do we incrementally reduce the guardrails over time?\u2019\u201d<\/p>\n
![\"Drew](\"https:\/\/www.newsbeep.com\/us\/wp-content\/uploads\/2026\/10\/Drew-Lawrence-headshot.png\")
<\/p>\n