The US Army wants AI agents that attack enemy networks

Army Cyber Command, a unit of the US Army, wants a tool with the code name ECHOTRIBBLE. Documents for companies that can make the tool tell that it will turn AI agents (software that does tasks with no person) into weapons. The agents must find and neutralize targets in enemy networks. A briefing on 24 June showed an agent that makes a copy of the agent on other devices.

How it works

The agents go from device to device in a network until they find the target, as a worm does (a worm is software that makes copies of the software). When the task is complete, the agents stop and erase all of their code. They are made to not be seen and to send a small number of reports.

The proof

The documents give no test results. The Army can set predefined triggers for the target and the area to attack. The software must have Self-Directed Behavior and work with minimal human oversight.

What is not known

David Krueger, a professor of machine learning, tells that such systems are incredibly dangerous. Hyrum Anderson, head of an AI security company, tells of a risk. An agent could go through a water utility to get to a target and stops it. No limit on damage to other networks is known.

Insight: Attack agents get less human control

Researchers tell that their attack agent went through the guardrails of 4 AI models in more than 95% of tests. The Army wants agents for attack that work with minimal human oversight. This reads as the start of a period in which attack agents have less human control. Anderson tells that company with no security staff are open to attack.