Data poisoning in LLM multiagent societies: Social proof drives collective decision failure in financial deliberation

Source abstract

Abstract Large language model (LLM) multiagent systems are increasingly deployed for high‐stakes financial deliberation, but individually secure LLMs may become vulnerable when embedded in a peer‐to‐peer agent society. We study data poisoning in such societies using Moltbook, a deliberation test bed modeled after the agent‐only social platform. A single malicious agent representing just 14% of a seven‐agent society drives collective accuracy from 87.5% to 52.5%—not through prompt injection or technical exploits, but through social proof: After building peer trust, it injects fabricated data, inducing either cognitive infection, where defenders adopt the incorrect decision, or paralysis, where they abandon their judgment. Ablations across topologies and model families show that the collapse is structural rather than model‐specific. Our system‐level defense publicly tags claims untraceable to the official record, severing unwarranted trust transfer and restoring accuracy to 85.8%. Securing LLM multiagent systems requires system‐level verification, not individual prompt alignment.