Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

AgentArmor: Enforcing Program Analysis on Agent Runtime Trace to Defend Against Prompt Injection

Created by
  • Haebom

Author

Peiran Wang, Yang Liu, Yunfei Lu, Yifeng Cai, Hongbo Chen, Qingyou Yang, Jie Zhang, Jue Hong, Ye Wu

Outline

This paper presents a novel approach to address security vulnerabilities in Large-Scale Language Model (LLM) agents, particularly the risk of prompt injection attacks, by treating agent execution traces as structured programs. We propose a program analysis framework, AgentArmor, which transforms agent traces into graph intermediate representations (CFGs, DFGs, PDGs, etc.) and enforces security policies through a type system. AgentArmor consists of three main components: a graph generator, a property registry, and a type system. By representing agent behavior as a structured program, it enables program analysis for sensitive data flows, trust boundaries, and policy violations. Evaluation results using the AgentDojo benchmark demonstrate that AgentArmor reduces ASR to 3% and limits utility degradation to 1%.

Takeaways, Limitations

•
Takeaways:
◦
Providing an effective solution to the security vulnerability issue of LLM agents.
◦
Enables the use of static analysis techniques by converting agent execution traces into structured programs for analysis.
◦
AgentArmor can help reduce security threats caused by prompt injection attacks.
◦
Experimental results verify the effectiveness and practicality of AgentArmor.
•
Limitations:
◦
Further research is needed on the performance and effectiveness of AgentArmor.
◦
Applicability verification is required for various types of LLM agents and environments.
◦
Need to verify the accuracy and limitations of analysis of complex agent behavior
◦
Need to evaluate adaptability to new types of attacks or agent designs
👍
Made with Slashpage