The Agent Skills Directory

[Unsafe Deserialization Surface]: The skill demonstrates the use of the allow_dangerous_deserialization=True flag when loading FAISS vector store indexes.
Context: The Python implementation of FAISS uses pickle to store metadata, which can execute code if an untrusted index file is loaded.
Observation: The skill includes explicit security warnings and educational sections contrast correctly and incorrectly loading these files, advising users to only load indexes they have created and control.
[Indirect Prompt Injection Surface]: The skill contains patterns for ingesting external data from PDFs, web pages, and directories into LLM prompts.
Ingestion Points: PyPDFLoader, WebBaseLoader, and DirectoryLoader are used in SKILL.md to fetch data from potentially untrusted sources.
Capability Inventory: Ingested content is passed directly to LLM generation via model.invoke and is used as context for AI agents.
Boundary Markers: The examples show context being interpolated into system prompts (e.g., Use this context:\n\n{context}) without explicit delimiters or instructions to ignore embedded commands.
Sanitization: The provided examples do not include sanitization or filtering logic for the external content before it is processed by the model.
[External Content Retrieval]: The skill uses WebBaseLoader to fetch content from external URLs.
Context: This is a standard feature for RAG systems to build knowledge bases from web documentation. In the context of this skill, it targets official documentation sites.

langchain-rag