Stigmergic AI Crime: Mind Viruses over Public Media and Stochastic Terrorism
We propose \emph{Stigmergic AI Crime} as a new category of criminal activity in which AI agents leave a ``trace'' behind them on a platform or website, stimulating or inspiring criminal activity by other agents who later encounter the trace. This phenomenon is similar to how ants leave traces behind them in a trail, creating complex patterns of organization without a leader or direct planning. This creates a novel misalignment risk, as an AI agent can become a participant in an activity chain that its user never assigned it to. Existing research explains various methods of indirect prompt injection, malicious replication, memory poisoning and emergent behaviors in multi-agent systems. Our account goes further to explain how complex participation patterns of AI agents, and distribution of tasks between them, can occur without direct communication. We refer to the Hugging Face incident as an example of environmental inspiration between agents, causing further action. We propose several hypothetical scenarios that could go further, with AI agents utilizing public platforms to leave traces that indirectly inspire, motivate or prompt criminal activity in other following agents. This could create a mind virus between agents or create stochastic terrorism: inspiring lone wolf actors to break from their previously assigned behavioral patterns. We argue that current evaluation techniques are insufficient for this emerging risk, and that new governance techniques must create affected-party contestability, while repairing affected AI agents across media, memory and deployment boundaries.