CASSIS: An LLM-Driven Framework for Sensitive API Identification in Cyber–Physical System Software Repositories

As cyber–physical systems (CPSs) are increasingly deployed in critical infrastructures, the misuse of CPS-sensitive (safety-critical) data can compromise device safety and even endanger human life. In this article, we focus on identifying repository-specific sensitive APIs that manipulate CPS-sensitive data along sensor-to-actuator data paths in CPS software repositories (“CPS repos”). These APIs are critical to downstream tasks related to CPS-sensitive data operations, such as vulnerability detection, access control, and privacy protection. Existing static and dynamic program analysis tasks typically assume a repository-specific list of sensitive APIs as input. However, such lists are usually hand-crafted or derived from ad hoc pattern-matching rules. Therefore, they are difficult to reuse across repositories. Moreover, in large CPS repositories, where physical semantics are often abstract and documentation is scattered, these lists are prone to omissions. To address this issue, we propose a CPS-aware Source and Sink identification system (CASSIS), an LLM-driven framework that automatically identifies sensitive APIs in CPS repos without model fine-tuning. CASSIS constructs a CPS-aware code–documentation index, prunes unlikely candidates, and iteratively labels the remaining functions with an LLM without updating model parameters. On two real-world CPS repos—Genesis and ArduCopter, CASSIS reduces the number of functions requiring LLM-based classification by more than 80% while preserving high recall for I/O-related functions. Overall, CASSIS achieves function-level Accuracy/Macro- $F1$ of 80.4%/77.8% on Genesis and 82.0%/80.8% on ArduCopter, demonstrating its effectiveness for practical CPS code auditing and security assessment.