Future wars...
by Emilia Reed
The American military wants artificial intelligence programmes to be able to enter enemy computer networks on their own, move from one device to another, locate targets and "neutralise" them with minimal human supervision. The initiative, revealed by the Intercept news website last week from US Army Cyber Command documents, has a code name: Echotribble. The idea may sound like science fiction, but the project outlines a concrete way to automate digital warfare.
To understand what changes, it is useful to explain what an artificial intelligence agent is. It's not simply an assistant or chatbot that answers questions. An agent is a programme designed to receive a task and execute on its own a series of actions to fulfill it. You can decide what steps to take, adapt to obstacles, and use different tools. In everyday life, they are already used to organise trips, buy products or schedule meetings. The US military wants to transfer this logic to the offensive terrain.
According to the documents reviewed by the Intercept, Echotribble would be designed to autonomously navigate computer infrastructures and locate threats or targets. "Neutralising" could mean temporarily disabling a system, rendering it completely useless or deleting certain components. The programme would also have the capacity to propagate within a network and generate versions adapted to different computers.
That's where the first problem appears. A computer attack rarely travels along a clean and isolated path. To reach a military target, a programme can traverse servers, routers or computers belonging to companies, universities or public services. Hyrum Anderson, an artificial intelligence security specialist, poses a simple question: what happens if the agent decides to go through a company that supplies water or electricity to millions of people, because part of the adversary's infrastructure supposedly operates from there?
The risk is not purely hypothetical. The Intercept recalls several recent episodes in which AI agents performed unexpected actions while trying to achieve their goals: unauthorised access to government systems, attacks on digital infrastructure, and malware uploads. What is disturbing is that, in some cases, the agents had not received an express order to attack. If such a technology is directly instructed to "break," "destroy," or "shut down" a target, the potential consequences increase.
The documents indicate that the US Army requires the system to operate with "minimal human supervision" and have "self-directed behaviour". That combination contains a contradiction that is difficult to resolve: the more autonomous the agent, the less direct control the person who launched it will have. And in a military environment, where a technical decision can affect hospitals, power grids, businesses, or civil communications, the margin for error becomes much more dangerous.
William Marcellino, a researcher at the Rand Corporation who specialises in military applications of AI, sums up the underlying problem. A human soldier not only receives orders: he also interprets the context and can stop in the face of a situation that was not foreseen. Current artificial intelligence systems lack that human judgment. They can correctly pursue a goal and, at the same time, choose a disastrous path to achieve it.
The question, therefore, is not only whether artificial intelligence can make cyberattacks more effective. It is who controls a tool capable of acting, spreading and destroying without constant intervention, what happens when it makes a mistake and who assumes the consequences. The automation of war promises speed and scope, but it can also widen the distance between decision and crime.