The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill demonstrates processing external content like images and PDFs, which represents an indirect prompt injection surface. Embedded instructions in these media files could potentially influence agent behavior. Ingestion points include data fields in contentBlocks and fs.readFileSync. No explicit boundary markers or sanitization techniques are provided in the examples. The primary capability utilized is model.invoke().\n- [DATA_EXPOSURE_AND_EXFILTRATION]: The examples use the fs module to access local files (such as ./photo.jpg and ./document.pdf). This local file access is limited to project-relative paths for the purpose of demonstrating multimodal data ingestion and does not involve sensitive system files or credentials.

langchain-multimodal