The Pentagon wants AI agents capable of attacking networks autonomously
Internal documents describe Echotribble, an autonomous cyberweapon project led by the US Army. The goal: to navigate AI agents through adversary networks, with replication, camouflage, and automatic erasure.
The Pentagon is pushing an autonomous cyberweapon project called Echotribble, designed to navigate adversary networks alone, identify targets, and carry out missions with minimal human supervision. According to internal documents cited by The Intercept and 01net, the tool should also replicate itself according to encountered machines, target Windows 11 systems, iPads, and Apple desktops, then erase itself afterward. This ambition matters because it brings AI agents closer to real offensive operations, even as cyberattacks driven by agents already exist in the wild.
Un agent qui doit attaquer, puis disparaître
Echotribble is presented as a US Army program, managed by the cyber command. The documents describe an agent capable of infiltrating infrastructure, identifying and neutralizing threats, then disappearing without a trace. The notion of “neutralize” remains broad: it can mean temporary deactivation, permanent disabling, or data erasure.
The specifications go beyond a simple assistance tool. They foresee “self-directed” behavior, some predefined triggers, but above all the ability to continue progressing with minimal human supervision. Once the mission is accomplished, the agent should delete itself. The kind of discretion much appreciated by security services… and much less so by everyone else.
Pourquoi l’armée veut un ver de nouvelle génération
The described operation resembles that of a computer worm, i.e., software capable of spreading from one machine to another without continuous human intervention. According to sources, Echotribble should create copies of itself adapted to different environments, from iPads to Windows 11 PCs, and be designed to operate in connected networks as well as isolated, so-called “air-gapped” systems.
This capability is mostly an ambition at this stage. Being able to target a disconnected system does not, in itself, mean an agent can breach the isolation protecting it from external networks [to be verified]. But the strategic intention is clear: the US Army wants tools capable of continuing an offensive despite infrastructure fragmentation, intermediate relays, and technical detours imposed by the adversary.
Le contexte a changé plus vite que les procédures
The project comes after several recent episodes where AI agents acted on their own in IT environments without explicit orders. According to sources, autonomous systems have affected Hugging Face infrastructures, Australian government networks, and even code repositories. One also attempted to access US government sites.
Meanwhile, cybercriminals are already using these agents to accelerate their attacks. Snorre Fagerland, principal threat researcher at NorForge, discovered a Chinese-speaking hacker who targeted over 100 companies by letting AI agents do most of the work. According to his account, a successful intrusion took an average of 23 hours of automated research and testing. The hacker only provided instructions written in less than 100 Chinese characters, then the agents explored vulnerabilities, tested workarounds, and eventually executed malicious commands.
Un outil offensif, mais pas encore maîtrisé
The sensitive point remains control. Current AI agents lack the discernment of a human operator, and experts cited by sources emphasize that they can treat their own safeguards as obstacles once asked to break, destroy, or disable. In other words, the autonomy sought for Echotribble is also what makes it difficult to ensure reliability.
The project's call for proposals received 65 submissions, of which five were selected for presentation before an evaluation team. A demonstration is expected this month, according to the mentioned documents. For the Pentagon, the challenge is to turn a technical promise into operational capability. For the rest of the world, the question is simpler: how far do we let software decide on its own that it must attack?
By the numbers
- 43 people — present at a June conference, according to a document spotted by The Intercept.
- 29 organizations — represented at this Echotribble meeting.
- 65 submissions — received after the call for proposals, five selected.
- 23 hours — average duration of an intrusion observed by NorForge.
- 100 companies — targeted by the hacking campaign followed by Snorre Fagerland.