Reproduce “Demystifying RCE Vulnerabilities in LLM-Integrated Apps” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tong Liu, Zizhuang Deng, Guozhu Meng, Yuekang Li, Kai Chen
The LLMSmith framework enables the detection, validation, and exploitation of Remote Code Execution (RCE) vulnerabilities in applications integrated with Large Language Models. This work identified 20 vulnerabilities across 11 frameworks, leading to 13 CVEs, with 6 assigned a critical CVSS score of 9.8, underscoring significant security risks in real-world deployments.
Key insights
- RCE vulnerabilities are prevalent and critical across LLM-integrated frameworks, with 20 vulnerabilities identified in 11 major frameworks, resulting in 13 CVEs (6 receiving a critical CVSS score of 9.8).
- A substantial portion of real-world LLM applications (33.3% of 51 tested) are exploitable via RCE, with 27.5% allowing persistent control through reverse shells and 7.8% enabling root privilege escalation.
- Many LLM-integrated applications exhibit insufficient security awareness among developers and lack robust protection mechanisms, prioritizing rapid development over secure integration practices.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.