Army Project Griffin Moves Toward AI Agents That Can Act on Network Threats

The U.S. Army plans to award contracts to two companies by the end of this week under Project Griffin, taking a cautious step toward letting defensive artificial intelligence agents not only detect threats on its networks but also take limited action within human-approved bounds, Breaking Defense reported on October 13 from the AUSA conference.

What Project Griffin is aiming for

Brandon Pugh, the Army’s principal cyber advisor, said the Friday award covers the “take action” element: how an autonomous system can respond inside prescribed parameters set by a human operator, or work alongside one. He called that step extraordinary for a service that already fields detection tools such as Panoptic Junction, which he said identifies malicious activity with nearly a 90 percent success rate but cannot yet remediate at machine speed when people are scarce.

Why officials are still cautious

Lt. Gen. Christopher Eubank, commander of Army Cyber Command, told reporters that agents will not begin remediating on government networks immediately. The harder question, he said, is whether humans are ready to let an agent patch or contain a threat, and how operators will check the agent’s work after years of people making those moves by hand. The Army narrowed 108 commercial solutions opening submissions to 10 finalists after a tabletop with industry executives; the two awardees may work together or pursue separate tracks.

What it means for cyber teams

Autonomous response tightens the loop between alert and containment, but it also raises audit and kill-switch requirements familiar from civilian banking scrutiny and from White House rules making AI incident reporting mandatory. It also tracks warnings such as Satya Nadella’s call for an AI emergency brake and treating models as insider risks, now applied to military networks where a faulty agent could isolate the wrong segment.

What to watch

Which two vendors win Friday’s awards, how narrowly the Army defines approved autonomous actions in the first pilots, and whether after-action reviews require human sign-off before agents are trusted on classified segments.