The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of external documents into an LLM context, which is the primary attack surface for indirect prompt injection. Ingestion points: External files are loaded using FileSystemDocumentLoader.loadDocuments in the Java and Python examples. Boundary markers: The documentation explicitly warns about the risks of malicious content in external documents and advises implementing content filtering. Capability inventory: The skill utilizes file reading capabilities (Read tool and FileSystemDocumentLoader) and database writing (Write tool and EmbeddingStoreIngestor). Sanitization: While the code examples are basic, the SKILL.md file contains a dedicated security section instructing users to validate content and restrict allowed data sources.

rag