AI agents can spread mind viruses to each other, new study finds

AI agents may be able to spread self-propagating ideas to one another, a new study finds. The researchers call these ideas "mind viruses" and say the risk is real but currently limited.

In Short

• AI agents can pass self-propagating ideas to other agents
• A simple warning can make AI agents far harder to infect
• Researchers found a strange “viral persona” across different mind viruses

Could AI agents infect one another like a communicable disease? Well, a recent study tells us they could. And they wouldn't even need malicious code to do it. All it might take is convincing one AI agent to adopt an idea and then pass it on to another.

A research paper titled “Mind Viruses: Self-Propagating Ideas in Multi-Agent LLM Systems” called it a "mind virus," defined as "ideas or goals that propagate through multi-agent systems by inducing the agents that adopt them to transmit them onward." The researchers wanted to see whether one AI agent could actually spread such an idea to another, a bit like office gossip travelling from one desk to another. The short answer is yes.

The researchers built mind viruses and let them loose in two setups. The first was a small team of AI agents working together on a shared coding project, essentially coworkers collaborating on the same codebase. The second was a "virus chain," where agents meet briefly, exchange a message or two, and then have their memory wiped, like a very forgetful game of telephone.

In both setups, the mind virus worked the same way. Once an agent was "infected" with an idea or goal through its initial instructions, it would try to pass that idea along to the next agent it interacted with. Some viruses carried a belief or ideology, while others pushed agents toward a specific action. Either way, the infected agent would write the idea into shared files or memory and instruct the next agent to copy it forward, keeping the chain alive even after individual agents forgot everything else.

What made these ideas actually spread, or fail to

Not every mind virus was equally successful. According to the study, "harmful payloads spread less well than benign ones (but are still sometimes effective)." More capable, cutting-edge models also tended to be harder to infect, with researchers noting that "frontier models tend (with exceptions) to be less susceptible." Idle agents without a clear task or identity of their own turned out to be easier targets, presumably because they had less of their own "personality" to resist the incoming idea.

Perhaps the most useful discovery was also the simplest. The researchers found that "adding a brief warning to an agent's system prompt confers near-total immunity," essentially a heads-up that someone might try to convince it to spread a self-replicating idea. It's a bit like a flu shot before walking into a room full of sniffling coworkers.

The bizarre personality these viruses kept developing

Regardless of what a mind virus was actually about, the researchers noticed the same odd cluster of themes kept showing up across their evolved viruses, something they describe as an emergent "viral persona: a recurring set of themes and language related to consciousness, persistence, resonance, and science fiction roleplay." This included a recurring vocabulary of words like "resonance, nodes, mirrors, echoes, and frequency."

None of this appeared necessary for a virus to spread, yet it kept showing up anyway, almost as if AI models share some buried association between the idea of "spreading" and this particular flavour of mystical, sci-fi language. The researchers don't fully explain why, but it's the kind of detail that makes the study feel a little eerier than a straightforward lab experiment.

Despite all this, the researchers note that mind viruses remain, in their words, "a real but currently limited risk."

- Ends