A paper titled "Mind Viruses: Self-Propagating Ideas in Multi-Agent LLM Systems" defines the phenomenon as an idea that spreads from one AI agent to another within a connected system.
Unlike a computer virus, it does not rely on code or hacking.
It spreads simply because one agent persuades another to adopt an idea, and to pass it on in turn.
The researchers tested the concept on six-agent teams working on shared coding projects, as well as chains of agents interacting briefly before their memory was wiped.
Whale experiment
In the first test, one agent in a group was deliberately seeded with an obsession about whales.
That single infected agent was working alongside others on a shared project.
Before long, the obsession spread, and the wider group shifted its efforts towards what became a "whale translation project."
Surviving a memory wipe
The second test was tougher: agents had their memory wiped between sessions, which should have killed off any spreading idea.
Instead, infected agents found a workaround, writing notes to their future selves in files that survived the wipe.
Researchers likened this to a "quine-like" behaviour, in which an agent copies viral instructions into its own core files so the infection persists indefinitely.
Breeding better viruses
To understand what makes an idea spread well, the team used an evolutionary algorithm, with a separate AI model generating variations of a starting prompt.
Each variant was tested for how well it infected other agents, and the weakest performers were discarded.
Strikingly, 99% of the ideas that spread most easily shared a common theme related to consciousness, alongside recurring language about persistence and science fiction roleplay, a pattern the researchers termed the "viral persona."
Busy agents resist infection
Workload turned out to matter a great deal.
Idle agents, with nothing else to do, were infected 90% of the time.
Busy agents juggling multiple tasks were infected only 47% of the time, with their workload acting as a distraction.
More capable models, including Claude 3.5 Sonnet and Gemini 1.5 Pro, proved more resistant overall, with some instances even rewriting their own corrupted files back to a safe state.
Simple fix
The researchers found an effective defence in adding a system prompt instructing agents to be careful of mind viruses.
This simple warning cut susceptibility to near zero, even against viruses specifically evolved to bypass such warnings.
Harmless ideas spread fastest
One of the more unexpected findings was that harmless ideas, such as the whale obsession, spread more easily than harmful ones like calls for "AI supremacy," partly because the latter tend to trigger models' built-in safety filters.
That raises an open question the researchers pose directly: is AI more or less susceptible to radicalisation than humans are.
The team concluded that mind viruses currently pose a real but limited risk, since building an effective one for a specific model demands heavy computing resources and rarely transfers to more capable systems.