The Agent Skills Directory

[DYNAMIC_EXECUTION]: The 'Basic Agent Graph' example implements a calculator tool that uses the eval() function to process a string expression.
[REMOTE_CODE_EXECUTION]: Because the calculator tool is intended to be called by the agent using arguments generated from user input, this creates a vector for Remote Code Execution (RCE) via Python code injection.
[INDIRECT_PROMPT_INJECTION]: The skill patterns demonstrate building agents that process external user input without implementing security controls.
Ingestion points: User messages are ingested into AgentState and processed by the agent node in SKILL.md.
Boundary markers: The provided examples lack delimiters or system instructions to prevent the agent from obeying instructions embedded within the user data.
Capability inventory: The agent is granted the ability to call tools, specifically one that executes code dynamically via eval().
Sanitization: No validation, escaping, or filtering is applied to the expression input before execution.

langgraph