The Agent Skills Directory

[DYNAMIC_EXECUTION]: The LangChain integration example in SKILL.md demonstrates the use of allow_dangerous_deserialization=True during the FAISS.load_local operation. This parameter allows the library to use Python's pickle module for loading index metadata. Loading a maliciously crafted index file from an untrusted source while this flag is enabled can lead to arbitrary code execution.
[INDIRECT_PROMPT_INJECTION]: The skill defines patterns for ingesting external data into vector indices, creating a surface for indirect prompt injection if the processed content contains adversarial instructions intended to influence downstream agent actions.
Ingestion points: Data is ingested via FAISS.from_documents in SKILL.md and index.add(vectors) in references/index_types.md.
Boundary markers: No specific delimiters or safety warnings regarding untrusted content are present in the code snippets.
Capability inventory: The skill utilizes file system operations for saving and loading indices (faiss.write_index, faiss.read_index, vectorstore.save_local, vectorstore.load_local).
Sanitization: The provided examples do not include sanitization, filtering, or validation of the input documents before they are processed for embedding.

faiss