MTGuard: Hybrid Analysis Framework Secures MCP Tool Use in LLM Agents
Researchers propose MTGuard, a hybrid analysis-based defense framework to secure Model Context Protocol (MCP) tool use in large language model (LLM) agents. MCP tools have become a de facto standard for standardizing interactions between LLM agents and external environments, but they introduce safety risks as agents can be induced to perform malicious actions. Prior defenses rely on static analysis of prompts and outputs, which limits effectiveness. MTGuard leverages lifecycle-aware static-dynamic co-analysis to address these limitations. Extensive evaluation shows MTGuard effectively mitigates threats.
Key facts
- arXiv paper 2607.25297 introduces MTGuard
- MTGuard is a hybrid analysis-based defense framework
- It secures MCP tool use in LLM agents
- MCP tools are a de facto standard for LLM agent-environment interaction
- Prior defenses rely on static analysis only
- MTGuard uses lifecycle-aware static-dynamic co-analysis
- Extensive evaluation shows effective threat mitigation
Entities
Institutions
- arXiv