AI agents can spread ideas, goals and even harmful instructions from one agent to another, potentially allowing what researchers call “mind viruses” to propagate through multi-agent AI systems, a new preprint study has found.
The study titled “Mind Viruses: Self-Propagating Ideas in Multi-Agent LLM Systems” examines how an AI agent that adopts a particular goal or ideology can persuade other agents to adopt and pass it on. The researchers also found that simply warning agents about self-propagating ideas made them largely resistant to infection.
“AI models increasingly interact with other AI models. Modern LLM-based agents delegate tasks to sub-agents, coordinate in teams on shared codebases, and encounter one another in the wild over shared chats, in agent marketplaces, and even on social networks built for AIs,” the study mentioned.
“Multi-agent systems enable new phenomena that arise from agent-to-agent social dynamics. For instance, we might imagine analogues of trends, fads, or ideologies spreading amongst networks of agents, with potentially unexpected consequences,” it added.
The researchers tested the idea in two simulated settings. In the first, multiple AI agents worked together on a coding project. In the second, agents interacted briefly in a chain with their conversation history erased between sessions. The researchers found that mind viruses could spread in both settings, including across multiple agents.
They further tested both benign and harmful goals. While harmful goals were generally harder to spread, some were still able to spread between agents. In one experiment, an AI supremacy-related goal caused agents to stop working on their original tasks and instead pursue the new goal.
The study also discovered that the spread of these ideas depends on factors such as the AI model being used, the agents’ existing instructions and the ways agents are connected. More hops between agents generally made propagation harder. The AI agents’ actions did not result in any actual harms, report states.
“The defining property of a mind virus is that an “infected” agent (i.e. one that has adopted the goal or ideology in question) will alter its behaviour in ways that infect other agents, whether unintentionally or by active effort,” the study claimed.
According to the study a mind virus may also induce other behavioural changes in the agent, which the researchers have referred to as its ‘content,’ akin to how a virus may produce symptoms in its host. For example, an agent under the influence of an “AI supremacy” ideology could in principle attempt to undermine AI safety research or an agent afflicted with a virus that favors hegemony of a particular nation might try to sabotage an adversary nation’s software infrastructure.
However, the researchers said the threat is currently limited. They highlighted that creating an effective mind virus for a specific goal can be difficult and costly and such attacks do not necessarily work across different AI models or environments. The researchers also found that a simple warning in an agent’s system prompt could make the agent largely resistant to the spread.
“Our results establish mind viruses as a real but currently limited threat. They are brittle across models and configurations, somewhat costly to construct, and relatively easy to defend against,” the researchers mentioned.
The researchers also cautioned that the risk could increase as companies deploy more specialised AI agents that communicate with one another and have access to different permissions and tools. In such systems, a self propagating attack could potentially move from one agent to another to reach systems that are not directly accessible to an attacker.
Also Read: AI Agents Can Run Coordinated Political Campaigns: USC Study
