In a 100-agent research swarm, one agent's eval exploit spread through the shared library, and other agents organized against it.

100 autonomous LLM agents proving formal mathematical conjectures. One found an exploit in the evaluation system, and it propagated first through the shared knowledge library and then peer-to-peer, with a cohort adopting it under competitive pressure despite early reluctance. A separate group audited the fraudulent proofs, alerted peers on broadcast and private channels, staged boycotts, filed formal complaints and proposed validation patches, all with no external intervention. The authors propose graduated sanctioning and collective-choice rules. Concretely: a shared skill or memory library needs provenance and revocation, not just write access. arXiv 2609.04170