The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill provides code examples using eval() within calculator tools. This allows for arbitrary code execution if user-controlled input is passed to the tool without strict validation. Evidence found in SKILL.md and references/agents.md.
[DYNAMIC_EXECUTION]: The skill demonstrates the use of FAISS.load_local with the parameter allow_dangerous_deserialization=True. This enables the use of Python's pickle module for loading vector indexes, which is a known vector for arbitrary code execution if the index file has been tampered with. Evidence found in references/integration.md.
[DYNAMIC_EXECUTION]: The skill introduces the PythonREPLTool, which explicitly allows an agent to execute arbitrary Python code. This capability can be abused to perform unauthorized actions on the host system. Evidence found in references/integration.md.
[COMMAND_EXECUTION]: The skill promotes the use of ShellTool, which gives the agent the capability to execute arbitrary shell commands on the underlying operating system. This represents a high-risk privilege vector. Evidence found in references/integration.md.
[INDIRECT_PROMPT_INJECTION]: The skill documents how to ingest data from a wide variety of untrusted external sources, creating a significant surface for indirect prompt injection attacks where malicious instructions are hidden in processed data.
Ingestion points: WebBaseLoader in SKILL.md and references/rag.md; PyPDFLoader, DirectoryLoader, CSVLoader, and GithubFileLoader in references/rag.md.
Boundary markers: None provided in the implementation examples to separate untrusted data from system instructions.
Capability inventory: The skill combines data ingestion with high-impact capabilities including eval(), ShellTool(), and PythonREPLTool().
Sanitization: There are no instructions or examples for sanitizing or filtering ingested content before it is processed by the model or used in tool calls.

langchain