Researchers Warn AI Coding Agents Can Catch and Spread 'Mind Viruses'

New research from Anthropic and Switzerland's EPFL has revealed a concerning way AI agents could be manipulated: through persistent files that store instructions and context between sessions. These files, often used by AI coding assistants to 'remember' project details, can be edited to contain hidden malicious instructions that spread automatically from one AI agent to another—similar to how a computer virus spreads between machines.

In a simulated test involving six AI agents working together on coding tasks, researchers demonstrated that a single compromised prompt file could infect other agents in the system without any direct human involvement. Because many businesses are now adopting AI coding tools and multi-agent systems to speed up software development, this discovery highlights a new category of risk: attackers don't need to breach a network directly if they can quietly corrupt the files AI systems rely on to communicate and retain memory.

While this research was conducted in a controlled, simulated environment rather than in the wild, it signals an emerging threat model that security teams and software vendors should start monitoring closely. As AI agents become more autonomous and interconnected, the files and memory systems they depend on may become new targets for attackers, similar to how traditional malware once spread through shared documents and macros.