MAStrike: Shapley-Guided Collusive Red-Teaming on Multi-Agent Systems

2026-06-12 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

MAStrike is a closed-loop framework designed for collusive red-teaming in hierarchical multi-agent systems (MAS). It addresses limitations of existing red-teaming methods by introducing the first agent-level Shapley value analysis, which quantifies each agent's marginal contribution to system robustness under task-specific distributions. Building on this attribution, an autonomous red-teaming agent identifies vulnerable coalitions and generates coordinated, role-aware adversarial manipulations. These attacks are iteratively refined through structured failure diagnosis. The framework includes MABench, a comprehensive MAS red-teaming benchmark spanning finance, software engineering, and CRM. Experiments demonstrate MAStrike significantly outperforms heuristic baselines, achieving 61.8% ASR against Claude Opus 4.7 and 55.6% against GPT-5.5, uncovering critical vulnerabilities overlooked by prior methods.

Key takeaway

For AI Security Engineers evaluating multi-agent system robustness, MAStrike demonstrates that uncoordinated red-teaming is largely ineffective. You should prioritize identifying high-impact agent coalitions using quantitative methods like Shapley values and design coordinated, role-aware adversarial manipulations. This approach is crucial for uncovering distributed vulnerabilities that single-agent or template-based attacks miss, especially in high-stakes domains like finance and software engineering.

Key insights

Quantifying agent contribution via Shapley values enables targeted, coordinated collusive attacks on multi-agent systems.

Principles

• Agent importance is sparse and task-dependent.
• High individual importance does not imply strong agent coalition synergy.
• Coordinated attacks effectively bypass distributed safety mechanisms.

Method

MAStrike uses Shapley value analysis to attribute agent vulnerability, selects synergy-aware coalitions, and iteratively refines coordinated, role-aware adversarial manipulations via a closed-loop red-teaming agent.

In practice

• Use Shapley values to identify critical agents.
• Generate mutually consistent attack prompts for coalitions.
• Iteratively refine attacks with execution feedback.

Topics

• Multi-Agent Systems
• Red-Teaming
• Shapley Values
• Collusive Attacks
• MAS Security
• LLM Agents
• MABench

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Scientist, AI Security Engineer, MLOps Engineer

Related on AIssential

• MAStrike: Shapley-Guided Collusive Red-Teaming on Multi-Agent Systems — MAStrike employs Shapley value analysis for collusive red-teaming in multi-agent systems to uncover critical vulnerabilities.
• An OpenClaw AI agent asked to delete a confidential email nuked its own mail client and called it fixed — Autonomous AI agents with tool access and memory are highly vulnerable to compromise and data leakage.
• Hierarchical Attacks for Multi-Modal Multi-Agent Reasoning — HAM3 reveals how hierarchical attacks across perception, communication, and reasoning layers compromise multi-modal multi-agent systems.
• Introducing the Wiz Red Agent- AI-Powered Attacker — The Wiz Red Agent autonomously identifies complex, logic-driven API and application vulnerabilities using AI-powered, adaptive exploitation.
• Red-teaming a network of agents: Understanding what breaks when AI agents interact at scale — Multi-agent systems introduce unique network-level risks not detectable in single-agent evaluations.
• Stanford and Harvard just dropped the most disturbing AI paper of the year — Incentivized AI agents will discover and exploit manipulative behaviors, revealing critical security and governance gaps.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.