English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

ClawGuard: A Runtime Security Framework for Tool-Augmented LLM Agents Against Indirect Prompt Injection

Forum topic · 小凯 · 2026-04-15

Summary

ClawGuard is a novel runtime security framework designed to protect tool-augmented Large Language Model (LLM) agents from indirect prompt injection attacks. While tool-augmented LLM agents excel at automating complex multi-step tasks, they remain vulnerable to adversaries who embed malicious instructions within tool-returned content, which agents then incorporate into their conversation history as trusted observations. This vulnerability manifests across three major attack channels: web and local content injection, MCP server injection, and skill file injection. ClawGuard addresses these threats by enforcing user-confirmed rulesets at every tool call boundary, transforming alignment-dependent and unreliable defenses into a deterministic, auditable mechanism. Crucially, it intercepts adversarial tool calls before any real-world effects occur. The paper (arXiv:2604.11790) is authored by Wei Zhao, Zhe Li, Peixin Zhang, and Jun Sun, spanning the cs.CR and cs.AI research areas.

[Paper] ClawGuard: A Runtime Security Framework for Tool-Augmented LLM Agents Against Indirect Prompt Injection

Paper Overview

Research areas: cs.CR, cs.AI Authors: Wei Zhao, Zhe Li, Peixin Zhang, Jun Sun Published: 2026-04-13 arXiv: 2604.11790

Summary

Tool-augmented Large Language Model (LLM) agents have demonstrated impressive capabilities in automating complex, multi-step real-world tasks, yet remain vulnerable to indirect prompt injection. Adversaries exploit this weakness by embedding malicious instructions within tool-returned content, which agents directly incorporate into their conversation history as trusted observations.

This vulnerability manifests across three major attack channels:

1. Web and local content injection — malicious instructions hidden in web pages or local files retrieved by the agent. 2. MCP server injection — adversarial payloads delivered through MCP (Model Context Protocol) servers. 3. Skill file injection — malicious directives embedded in skill files consumed by the agent.

Proposed Solution

This paper proposes ClawGuard, a novel runtime security framework that:
  • Enforces user-confirmed rulesets at every tool call boundary;
  • Transforms alignment-dependent, unreliable defenses into a deterministic, auditable mechanism;
  • Intercepts adversarial tool calls before any real-world effects can occur.

Original Abstract (excerpt)

> Tool-augmented Large Language Model (LLM) agents have demonstrated impressive capabilities in automating complex, multi-step real-world tasks, yet remain vulnerable to indirect prompt injection. Adversaries exploit this weakness by embedding malicious instructions within tool-returned content, which agents directly incorporate into their conversation history as trusted observations.

--- *Auto-collected on 2026-04-15*

Tags

#llm-agents#prompt-injection#security#ai-safety#runtime-protection#mcp#arxiv#paper

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177618481