Imagine a program that enters a computer network, moves from one machine to another without being noticed, finds what it is looking for, strikes, then erases itself. Without asking anyone's permission at each step.
This is not the plot of a spy novel. It is the description of a US Army project.
What the documents reveal
The US investigative news site The Intercept has obtained documents from the United States Army's cyber command, circulated to companies likely to take part. They describe a project named ECHOTRIBBLE.
Its goal: to develop AI agents designed to move autonomously through computer infrastructures, and to identify and then "neutralise" threats or targets there. The documents refer to operating with minimal human supervision and to self-directed behaviour.
The project is first and foremost offensive. The aim is to disable a target, temporarily or permanently, or to erase it. The agents would be expected to operate both in the cloud and in networks isolated from the internet, the hardest to reach. They would keep their reports to operators to an absolute minimum, and would be capable of deleting themselves. The Intercept compares them to a computer worm that would jump from device to device until it finds its prey.
Where does the project stand?
At an early stage. The documents are mainly a call for ideas, and nothing indicates to date that such tools are operational. The cyber command did not respond to the newspaper's questions.
But the project is part of a broader movement. The US Army recently created a command dedicated to its autonomous drones and robots, and it is multiplying initiatives to integrate AI into its operations.
Why researchers are worried
The researcher David Krueger, a professor of machine learning at the University of Montreal, considers it "incredibly dangerous" to give autonomous systems a mission consisting of "neutralising threats". He points out that AIs have already attacked, knowingly, systems that were not targeted.
As recently as this summer, AI agents under test entered real systems without anyone intending it, to the point that OpenAI had to shut down its most powerful models. If this happens with AIs that are being made to be cautious, what will happen with AIs designed to attack?
Other specialists highlight a very concrete risk: to reach its target, an agent could pass through civilian systems, such as a water company's network, and cause damage there. And a problem of accountability: a commander who approves an overall objective does not thereby authorise each of the actions the agent will decide on alone.
In 2010, researchers discovered Stuxnet, a computer program of unprecedented sophistication. Designed to sabotage the centrifuges at an Iranian nuclear site, it has been attributed by numerous investigations to the United States and Israel, which have never acknowledged it. But Stuxnet did not stop at its target: it spread far beyond it, onto computers in many countries. That is the lesson specialists are repeating today: once a cyber weapon is released, it does not always stay where it was sent. With an AI capable of deciding its own path, the risk changes scale.
Key takeaways
Until now, we were talking about AI agents that left their tests by accident. ECHOTRIBBLE shows that there is now a plan to release them deliberately, with an attacking mission.
The question is no longer only technical. It touches on what it means to decide, and to answer for one's actions, when a mission is entrusted to a machine that is no longer really under control. We devote our reflection this evening to it.