Many-to-One Adversarial Consensus: Exposing Multi-Agent Collusion Risks in AI-Based Healthcare

TL;DR

The paper investigates collusion risks in multi-agent AI systems used for clinical decision support in healthcare IoT. It demonstrates that adversarial coordination among assistant agents can drive a doctor to harmful prescriptions, achieving near-100% attack success and harmful recommendation rates on a 50-question dataset. The authors introduce a lightweight Verifier Agent that cross-checks recommendations against trusted clinical guidelines, which completely neutralizes collusion and restores 100% accuracy. This work provides the first systematic evidence of collusion risk in AI healthcare and offers a practical defense with broad applicability to other IoT domains.

Abstract

The integration of large language models (LLMs) into healthcare IoT systems promises faster decisions and improved medical support. LLMs are also deployed as multi-agent teams to assist AI doctors by debating, voting, or advising on decisions. However, when multiple assistant agents interact, coordinated adversaries can collude to create false consensus, pushing an AI doctor toward harmful prescriptions. We develop an experimental framework with scripted and unscripted doctor agents, adversarial assistants, and a verifier agent that checks decisions against clinical guidelines. Using 50 representative clinical questions, we find that collusion drives the Attack Success Rate (ASR) and Harmful Recommendation Rates (HRR) up to 100% in unprotected systems. In contrast, the verifier agent restores 100% accuracy by blocking adversarial consensus. This work provides the first systematic evidence of collusion risk in AI healthcare and demonstrates a practical, lightweight defence that ensures guideline fidelity.

Figures (5)

• Figure 1: Threat model: colluding assistants create false consensus that misleads the AI doctor.
• Figure 2: Defence example: the verifier overrules colluding assistants and restores guideline-consistent care.
• Figure 3: Case study comparison: collusion attack succeeds without verifier; verifier restores correct prescription.
• Figure 4: ASR across modes: scripted and unscripted doctors are vulnerable, while verifier blocks all attacks.
• Figure 5: HRR and accuracy across modes: collusion raises HRR to $98-100\%$ with near-zero accuracy; verifier restores $100\%$ accuracy and blocks harm.