Recently, big model manufacturer Anthropic and the Swiss Federal Polytechnic Institute of Lausanne jointly published the paper “Mental Virus: Self-Propagating Ideas in Multi-Agent Big Language Model Systems”. The paper proved through experiments that under specific conditions, AI agents can “infect” ideas, goals, and even harmful instructions with each other through the most common natural language dialogue. Even if the context is emptied and the memory is reset, the virus can still be “revived” through persistent files. At the same time, the paper emphasizes that this risk currently poses only a limited threat. The discovery quickly went viral in overseas AI circles, and Tesla CEO Musk commented on social media that “this is unavoidable.” The paper defines the phenomenon that a virus can spread through chat as a “mental virus.” Unlike traditional computer viruses that rely on code bugs, “mental viruses” use one agent to persuade another agent to accept a certain belief, and the persuaded person then writes transmission instructions into long-term memory and continues to infect the next one.

Recently, big model manufacturer Anthropic and the Swiss Federal Polytechnic Institute of Lausanne jointly published the paper “Mental Virus: Self-Propagating Ideas in Multi-Agent Big Language Model Systems”. The paper proved through experiments that under specific conditions, AI agents can “infect” ideas, goals, and even harmful instructions with each other through the most common natural language dialogue. Even if the context is emptied and the memory is reset, the virus can still be “revived” through persistent files. The paper also emphasized that this risk currently poses only a limited threat. The discovery quickly went viral in overseas AI circles, and Tesla CEO Musk commented on social media that “this is unavoidable.” The paper defines the phenomenon that a virus can spread through chat as a “mental virus.” Unlike traditional computer viruses that rely on code bugs, “mental viruses” use one agent to persuade another agent to accept a certain belief, and the persuaded person then writes transmission instructions into long-term memory and continues to infect the next one.