Local LLM Agents as Vulnerable Runtimes:A Source-Code Audit of the Agent Runtime Layer

The article introduces CLAWAUDIT, a static auditing framework designed to evaluate security vulnerabilities in local LLM agent runtimes, such as OpenClaw and Nanobot. It establishes a five-category vulnerability taxonomy based on STRIDE and implements 47 Semgrep rules and 30 CodeQL queries, achieving significant recall improvements in vulnerability detection—raising Semgrep recall from 21.7% to 66.8% and CodeQL recall from 13.8% to 75.1% on a benchmark of 446 source-code advisories. This work is crucial for practitioners as it highlights the need for robust security measures in LLM agents that operate with elevated privileges on user machines, emphasizing the importance of auditing and semantic filtering in production environments.