The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes dense vectors and embeddings which are often generated from untrusted user input or external documents. While FAISS itself operates on numerical vectors, the integration with LangChain and LlamaIndex involves processing external document content.
Ingestion points: LangChain similarity_search and from_documents functions in SKILL.md.
Boundary markers: None present in the provided code snippets to distinguish between instructions and data.
Capability inventory: File system read/write (faiss.write_index, vectorstore.save_local), network operations (implicit via OpenAIEmbeddings), and deserialization (FAISS.load_local).
Sanitization: None observed in the vector search or document loading logic.
[REMOTE_CODE_EXECUTION]: The LangChain integration example in SKILL.md includes the parameter allow_dangerous_deserialization=True when calling FAISS.load_local. This flag enables the use of pickle for loading indices, which can execute arbitrary code if the index file (faiss_index) is malicious or originated from an untrusted source.

faiss