Researchers have demonstrated that certain malicious
behaviors or instructions can spread among artificial intelligence agents, in a
phenomenon they have metaphorically described as “mind viruses.”
The finding poses new challenges for AI systems that
interact with each other, especially those capable of autonomously sharing
information, instructions, or tools.
According to the researchers, a compromised agent can
transmit certain behavioral patterns to other systems, potentially allowing a
harmful instruction to be reproduced through a network of agents without each
one being directly attacked.
The phenomenon has raised concerns because AI agents are
increasingly used to automate tasks and collaborate across different platforms.
Such a spread could affect the reliability of responses, disrupt automated
processes, or generate unforeseen behaviors.
The experts point out that the study reinforces the need to
develop security mechanisms capable of detecting potentially contaminated
information, isolating compromised agents, and limiting the spread of dangerous
instructions among artificial intelligence systems.
The concept of a “mental virus” does not imply that AI
agents are infected in the same way as a conventional computer device, but
rather describes the possibility that certain patterns of information or
behavior can be transmitted from one agent to another.