The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill demonstrates creating a 'Calculator' tool using the Python eval() function on LLM-generated strings (SKILL.md, references/agents.md). This allows for arbitrary code execution if the agent is manipulated into generating malicious expressions.
[DYNAMIC_EXECUTION]: The documentation includes the use of PythonREPLTool and ShellTool (references/integration.md), which grant the agent the ability to execute arbitrary Python code and shell commands on the host system.
[DYNAMIC_EXECUTION]: The FAISS integration guide (references/integration.md) shows the use of allow_dangerous_deserialization=True, which enables the loading of potentially malicious pickle files during vector store loading.
[INDIRECT_PROMPT_INJECTION]: The skill focuses on RAG (Retrieval-Augmented Generation) and agents that ingest data from external sources like web pages, PDFs, and GitHub repositories. Ingestion points: WebBaseLoader, PyPDFLoader, GithubFileLoader, and DirectoryLoader are used to pull content into the agent's context. Boundary markers: Examples use basic prompt templates (e.g., 'Context: {context}') without robust delimiters or instructions to ignore embedded commands. Capability inventory: The agent is equipped with powerful capabilities including eval(), PythonREPLTool, ShellTool, and SQLDatabase access. Sanitization: There are no examples of sanitizing or validating the retrieved external content before passing it to the LLM or tools.

langchain