Automated cross-modal detection that jointly analyzes natural language and code

Develop automated detection methods that jointly analyze natural-language skill descriptions and executable source code in LLM agent skills to identify credential leakage arising from cross-modal interactions.

Background

The analysis shows that 76.3% of credential leakage cases require joint consideration of natural language instructions and program logic; neither modality alone suffices to reveal the exposure.

Existing secret detection approaches focus on single-modality analysis, and the authors explicitly identify the need for automated detection that jointly analyzes natural language and code as an open problem.