Neo: Detecting Privilege Escalation Vulnerabilities in Multi-Language Microservices with Agentic Program Analysis

Researchers propose Neo, an agentic program analysis framework that combines LLMs with classical program analysis to detect privilege escalation vulnerabilities in microservices, finding 24 zero-day vulnerabilities in 25 open-source applications. Its LLM agent dynamically generates analysis plans and adjusts code search strategies; the evaluation covers 7 programming languages and 6.2 million lines of code. Neo achieves 81.0% precision and 85.0% recall on a benchmark dataset, and applying it to other domains and vulnerability types found 18 additional zero-day vulnerabilities.