The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements Retrieval Augmented Generation (RAG) which involves ingesting and processing untrusted external data (documents) and interpolating them into LLM prompts.
Ingestion points: Data enters the system via the /documents POST endpoint in examples/rag-implementation.js and is later retrieved and injected into prompts in the queryKnowledgeBase function.
Boundary markers: The implementation uses clear structural separation in the prompt (e.g., Context:\n${context}\n\nQuestion: ${question}) and includes a system instruction to 'Answer the question using ONLY the information provided in the context.'
Capability inventory: The skill uses env.AI.run for inference and embeddings, env.DB (D1) for storage, and env.VECTOR_INDEX (Vectorize) for similarity search. It does not perform arbitrary shell command execution or external network requests beyond Cloudflare's platform services.
Sanitization: The skill uses structured JSON for request handling but lacks explicit sanitization or filtering for malicious instructions embedded within the ingested text chunks beyond the instructions provided to the LLM.

Workers AI