Abstract
AI agents increasingly rely on tools built on or dependent on open source software (OSS) to access data, invoke services, and modify external resources. Vulnerabilities in these dependencies can there- fore propagate into agent systems. Conventional analyses trace vulnerability effects through explicit program states and data or control dependencies. Agent systems introduce model inference be- tween tool output and subsequent action, so affected states become model-visible observations whose influence depends on semantic interpretation. Program analysis can show that such values reach model input, but not whether they change the generated action or external system. We assess three successive links in this process. Vulnerable execution must change the observation, the changed observation must influence the generated action, and the executed action must change an external resource. Our framework combines program tracing, repeated executions under matched contexts, and external state verification to establish these links.
Funding Information
  • Major Technological Innovation Program of HUST (2025ZDKJCX05)