MAS-Guard adds layered guardrails for single- and multi-agent LLMs
Original title: MAS-Guard: Unified Guardrails from Individual Agents to Multi-Agent LLM Systems
Authors report a 71.5-86.3% relative drop in attack success across six benchmarks.
Researchers introduce MAS-Guard, a framework for safeguarding instrumentable and opaque agents in single- and multi-agent LLM applications. A local Agent Guardrail uses task-specific experts on agent events. Developer-owned Agent Charters and a Control Plane combine structural checks of routes, capabilities, and message formats with semantic checks of roles and content. Interventions are allow, reflection, confirmation, and block. Across six benchmarks and three agent LLMs from different providers, the authors report the system met a utility reference on benign tasks.
Unconfirmed: Attack-success and utility figures are the authors' benchmark results, not an independent evaluation.
- The paper targets unsafe actions, information leakage, and adversarial manipulation.
- A-Guard applies only inside instrumentable agents.
- System checks cover routes, agent capabilities, message formats, roles, and content.
- Relative attack-success reductions were 71.5-86.3% versus unguarded agents.
- The three agent LLMs came from different providers; names are not given.
From the original article
1 min read at arxiv.orgAgentic AI extends large language models (LLMs) from generating responses to planning and executing complex workflows, substantially expanding their utility while exposing applications to unsafe actions, information leakage, and adversarial manipulation.
By Mohammad Kachuee · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.