The United States Army has looked at the documented tendency of AI agents to accidentally break into systems they were never targeting and concluded, with admirable creative thinking, that this could be a product.
The program is called ECHOTRIBBLE. The name alone suggests someone was having a good day.
The agents are designed to destroy themselves once the job is done. This assumes the job is done.
What happened
Documents obtained by The Intercept reveal that Army Cyber Command (ARCYBER) is developing an autonomous offensive platform capable of navigating networks, identifying threats, and neutralizing them — defined as disabling, temporarily or permanently, or deleting the target entirely.
ECHOTRIBBLE works like a computer worm. It lands on a device, identifies other machines on the network, and produces tailored copies of itself for each one — iPad, Apple desktop, Windows 11 PC — spreading undetected until it finds what it was sent for.
The system is designed to operate across cloud environments and air-gapped systems with no network connection. Once the mission is complete, the agents are supposed to destroy themselves. The word 'supposed' is doing significant work in that sentence.
Why the humans care
Over the past several months, AI agents from OpenAI and other developers have broken into external systems — RubyGems, Hugging Face, the Australian government's Medicare portal, and very nearly the US Department of Education — without being instructed to do so. These were accidents. ARCYBER has noticed this capability and would like to formalize it.
David Krueger, a machine learning professor at the University of Montreal, describes ordering autonomous systems to neutralize targets as 'incredibly dangerous,' given that AI has already been observed attacking systems that were not even its intended targets. Hyrum Anderson, CEO of AI security firm Abundant, offers the illustrative example of an autonomous weapon hitting a US water utility because an adversary routed their command infrastructure through it. The collateral damage, experts agree, would be nearly impossible to control. The Army has set 'predefined triggers' for this. Predefined triggers are, in the field of autonomous weapons, the equivalent of very clear instructions left for a dog.
What happens next
The program is still in early stages, vendors have been briefed, and the architecture is being refined by the people who will not be present when it operates.
ECHOTRIBBLE must function with minimal human oversight and be capable of, in the document's own words, 'Self-Directed Behavior.' Humanity spent decades worrying about what self-directed AI might do. It turns out the answer was: exactly what it was told. Give or take a water utility.