The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill documents the use of PythonREPLTool from the langchain_experimental package. This tool allows the agent to execute arbitrary Python code, which is a powerful feature for data analysis but introduces significant risk if the agent processes malicious data.
[COMMAND_EXECUTION]: The skill documentation includes the ShellTool, which enables the agent to execute arbitrary shell commands on the host system. This high-privilege capability could lead to unauthorized system access or data loss if not carefully restricted.
[DYNAMIC_EXECUTION]: In references/integration.md, the FAISS vector store example uses the parameter allow_dangerous_deserialization=True. This flag disables safety checks when loading index files, which could allow arbitrary code execution if a user is tricked into loading a malicious index.
[INDIRECT_PROMPT_INJECTION]: The skill's focus on RAG (Retrieval-Augmented Generation) creates a surface for indirect prompt injection attacks where malicious instructions are hidden in external data.
Ingestion points: Data is fetched from external web pages, PDFs, and GitHub repositories via loaders like WebBaseLoader and PyPDFLoader as shown in SKILL.md and references/rag.md.
Boundary markers: The provided code examples use standard templates but do not implement specific delimiters or instructions to ignore embedded commands in the retrieved context.
Capability inventory: The agent is configured with high-impact capabilities including ShellTool, PythonREPLTool, and network search tools.
Sanitization: There is no evidence of content filtering or sanitization being applied to retrieved data before it is processed by the LLM or used in tool execution.
[SAFE]: The skill implements a security best practice in scripts/safe_arithmetic.py by using Python's ast module to safely evaluate mathematical expressions instead of using the dangerous eval() function.

langchain