The Agent Skills Directory

[COMMAND_EXECUTION]: The code example for the calculator tool in SKILL.md implements a math evaluator using the Python eval() function. This is a significant security risk because eval() executes the string passed to it as Python code. If an attacker provides a malicious string through the agent's input, they could execute arbitrary shell commands on the host environment.
[REMOTE_CODE_EXECUTION]: The use of eval() within a tool definition represents a remote code execution vulnerability. In an agentic workflow where an LLM determines the arguments for a tool, an indirect prompt injection or a direct malicious query could lead to the execution of arbitrary Python code, potentially compromising the entire application and its environment.

langgraph