US Army awards AI agents that counter cyberattacks: the bottleneck is human trust

By National Defense Magazine · October 12, 2026.

The U.S. Army will award two contracts this week for Project Griffin, the first of three AI projects for its cyber defense, service officials announced on October 12 at the annual meeting of the Association of the United States Army (AUSA). The solicitation was open for just one week and drew 108 proposals; the Army narrowed them down to 10 companies and will pick two. According to Brandon Pugh, principal cyber adviser to the Secretary of the Army, the idea behind having two vendors is that they can interact with each other or keep operating independently.

The program, called Army Rapid Development of Cyber Defense Systems, grew out of a tabletop exercise with 14 executives from the defense industry and multinationals already using AI in cyber missions. Some 20 possibilities emerged, and the Army chose three lines of effort to narrow the scope. The premise, Pugh said, is what can be done if thousands of autonomous attacks are launched against the Army.

Griffin aims to move from detection to action. The Army already uses an AI-powered monitoring platform, Panoptic Junction, which Pugh said detects malicious activity with roughly 90% accuracy. The goal is that, upon such a detection, a human operator won't have to jump on the keyboard immediately: the agent would carry out a series of response actions within guardrails set by the Army. It will start with the Army's portion of the Department of Defense information network and expand as lessons are learned.

The other two projects are agentic deception and "purple teaming." The former would try to lure attackers away from the "crown jewels" toward areas with less important data, so they spend "their tokens" (the unit of compute for language models) far from what matters. The latter would automate defense testing, a process that has so far been highly manual. Pugh gave no dates for these solicitations but said he is pushing to shorten them.

One detail is telling: the Army doesn't yet know what specific actions the agents will take. Lt. Gen. Christopher Eubank, head of Army Cyber Command, said they are talking with industry to understand what that "action space" means and what limits it requires. Pugh added that the call was left broad at the companies' request, as they didn't want to be boxed into two or three capabilities. The plan is iterative: start with a few actions, test them on its network and expand.

**Our take.** The most important line in the story isn't technical. Eubank summed it up: the question is whether humans are ready for the agent to act, after years in which a person applied the patch or the remediation, and how the agent's work gets checked. It's the same conclusion much of the industry is pointing to: the limit on agents is no longer model capability, but governance over what they can touch and decide without supervision.

It makes sense. If the attacker automates, the human defender loses on speed: an incident resolved in minutes is already too late against thousands of simultaneous attempts. But granting an agent permission to act on a military network introduces a new risk: an agent that makes a mistake, or is manipulated, can cause damage from the inside. That's why it's sensible to start with a few actions, in a contained environment, and measure. The 90% accuracy figure also deserves a caveat: it comes from the institution itself, with no public detail on false positives, and in automated defense a false positive is no longer an annoying alert but an executed action.

There's also a message about procurement. Opening a solicitation for one week and evaluating 108 proposals is a real change from months-long cycles. Pugh wants to adapt commercial solutions "in weeks, not years." That favors those who already have a product, and reinforces an idea we keep tracking: value is shifting from the model to the layer of permissions, auditing and control between the agent and the systems.

In the short term, autonomous cyber defense will bring errors, escalations and uncomfortable debates about accountability. In the long term, it's hard to imagine secure critical infrastructure without defenses operating at machine speed. The key will be how much is invested in verifying the agent, not just in building it.