The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The RAG implementation in SKILL.md creates a vulnerability surface by interpolating untrusted data from the vector database and user queries directly into a large language model prompt without sanitization or boundary markers.
Ingestion points: The context variable (populated by doc.content from the vector store) and the question parameter in the RAGService.query method.
Boundary markers: Absent. The prompt template uses simple string concatenation without delimiters or instructions to treat the injected content as data rather than instructions.
Capability inventory: The skill allows high-privilege tools including Bash, Write, and Edit, which could be targeted if a malicious document in the vector store successfully triggers an injection.
Sanitization: The code lacks any escaping, filtering, or validation of external content before interpolation.

vectordb