Evaluating LLM Agents on Automated Software Analysis Tasks

Abstract

Numerous software analysis tools exist today, yet applying them to diverse open-source projects remains challenging due to environment setup, dependency resolution, and analysis tool configuration. LLM-based agents offer a potential solution, yet their effectiveness on the specific task of automated software analysis has not been systematically studied. Unlike issue solving or general environment setup, this task requires installing and configuring a separate analysis tool alongside the target project, generating tool-specific prerequisites, and validating that the analysis tool produces meaningful analysis outputs rather than merely declaring successful termination without evidence. We introduce AnalysisBench, a benchmark of 35 tool-project pairs spanning seven analysis tools and ten diverse C/C++ and Java projects, each with a manually constructed reference setup. Using AnalysisBench, we evaluate four agent architectures across four LLM backends. Our custom agent, AnalysisAgent, achieves manually verified success rates of 94% (Gemini-3-Flash, 33/35 tasks), compared to 77% for the best baseline (ExecutionAgent). Beyond quantitative results, we identify key limitations in existing agents, including stage mixing, poor error localization, and premature termination, and show that agentic architecture plays a critical role beyond LLM capability alone. We further find that whole-program analyses and Java-specific tools are the most difficult tasks, that Java toolchains pose greater challenges than C/C++, and that self-validated success consistently overstates manually verified success. Extended runs with AnalysisAgent-produced setups also surface two previously unknown defects in masscan and radare2, confirming that the produced setups support real downstream analysis.