The US Army plans to award contracts to two companies to develop AI agents that could respond to cyber threats on military networks. The agents are not about to be let loose on live systems: Army officials say deployment will follow a deliberate testing process, with human oversight still a central question.
Watch Desk analysis
What happened
Breaking Defense reports that Brandon Pugh, the Army’s principal cyber adviser, said at the AUSA conference in Washington that the contracts are due to be awarded by Friday. The work is part of Project Griffin, an Army effort to develop defensive AI agents for its networks.
The Army received 108 submissions to a commercial-solutions call and narrowed them to 10 finalists, according to the report. Pugh said the work is intended to move beyond detecting a potential adversary towards taking action against a threat within prescribed parameters, either autonomously or alongside a human operator. The two companies may work together or separately.
The report says the agents will not begin work on government networks immediately. Lt Gen Christopher Eubank, commander of Army Cyber Command, said the Army is working with the companies to understand what the agents can do and how to check their work before any deployment.
Why it matters
This is a shift from asking AI to spot a threat to considering whether it should help remediate one. That could matter to network defenders facing more alerts than people can readily handle. It also raises the harder operational question: what should an agent be allowed to change, and how does a human know its intervention was sound?
The Army already has tools it uses to identify malicious activity, Pugh told Breaking Defense, but he said it lacks a way to remediate threats, particularly with limited personnel. The contracts are a step towards testing a possible response, not evidence that autonomous action is already operating on Army networks.
Our read
Defensive agents could be useful when the alternative is leaving a threat unaddressed while a stretched team works through a queue. But “within prescribed parameters” is doing important work here. The real test is whether those boundaries can be made dependable, and whether operators can understand and review what the agent does.
What to watch
- Which two companies receive the contracts and whether they work together.
- What actions the agents will be permitted to take during testing.
- How the Army plans to check agent decisions and intervene when needed.
- Whether the effort progresses from development into trials on government networks.
Discussion spark: Should a defensive AI agent be allowed to remediate a cyber threat on its own within set limits, or should a human approve every action?
Sources and evidence
- Source update (13 October 2026, 11:00 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.