I Can't Believe It's Corrupted: Evaluating Corruption in Multi-Agent Governance Systems

Researchers evaluated multi-agent governance simulations in which large language model agents held formal government roles under different authority structures, scoring violation and abuse outcomes across 28,112 conversation records with independent rubric-based review. Results showed that, among models that had not reached saturation, governance structure drove corruption-related outcomes more than model identity, with large differences across governance regimes and model-governance pairings; lightweight guardrails reduced risk in some settings but did not reliably prevent severe failures. The authors accordingly argue that the integrity of institutional AI should be a pre-deployment requirement rather than a post-deployment assumption, and that before granting LLM agents real authority they should be stress-tested under governance-like constraints, including enforceable rules, auditable logs, and human oversight of high-impact actions.