The Agent Skills Directory

Unsafe Deserialization Consideration: The skill demonstrates using allow_dangerous_deserialization=True when loading local FAISS indices in SKILL.md. This pattern uses Python's pickle module under the hood, which can execute arbitrary code if the index file being loaded has been tampered with or originates from an untrusted source.
Indirect Prompt Injection Surface: The skill ingests data from external sources (PDFs, local directories, and websites) and interpolates this content directly into the LLM prompt.
Ingestion points: PyPDFLoader, WebBaseLoader, DirectoryLoader, and the search_docs tool in SKILL.md.
Boundary markers: The skill uses basic string interpolation (Use this context:\n\n{context}) without robust delimiters or explicit instructions to ignore embedded commands.
Capability inventory: The agent is provided with a search_docs tool that can fetch potentially attacker-controlled content.
Sanitization: There is no evidence of sanitization or filtering applied to the retrieved content before it is processed by the agent.
External Data Fetching: The skill utilizes WebBaseLoader to retrieve content from external URLs (specifically docs.langchain.com). While fetching from documentation is a standard RAG use case, users should be aware of the network operations performed during document loading.

langchain-rag