The Agent Skills Directory

[Indirect Prompt Injection] (LOW): The RAG chatbot template in templates/rag-chat.ts is vulnerable to indirect prompt injection. It retrieves documents from a vector database and interpolates their content directly into the system prompt for the LLM.
Ingestion points: User input question and vector database matches are ingested in templates/rag-chat.ts (lines 52 and 81).
Boundary markers: Minimal string separators (---) are used, which are insufficient to prevent an LLM from following instructions embedded within the retrieved data.
Capability inventory: The script uses env.AI.run to interact with the LLM.
Sanitization: No input sanitization or escaping is performed on the context data before prompt interpolation.
[Data Exposure & Exfiltration] (SAFE): No hardcoded credentials or sensitive file path access detected. The skill correctly utilizes Cloudflare environment bindings for database and AI services.
[Obfuscation] (SAFE): No obfuscation, hidden Unicode characters, or encoded commands were found in the templates or documentation.
[Unverifiable Dependencies] (SAFE): No external script downloads or third-party package installations are performed by the skill's code. Documentation refers to standard, official Cloudflare tools like wrangler.

cloudflare-vectorize