The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill provides multiple implementation examples for 'ReAct' agents and tool use that utilize the Python 'eval()' function to process logic for a calculator tool. Since the input to this tool is generated by an LLM based on user queries, this creates a significant security risk where an attacker could execute arbitrary code by manipulating the agent's reasoning process. Evidence found in 'SKILL.md' and 'references/examples.md'.
[COMMAND_EXECUTION]: The use of 'eval()' in the agent tool examples facilitates the execution of arbitrary Python commands within the host environment.
[INDIRECT_PROMPT_INJECTION]: The skill documents patterns for agents that process untrusted user data, which can lead to indirect prompt injection. Ingestion points: The 'question' parameter in 'react_agent' functions across 'SKILL.md' and 'references/examples.md'. Boundary markers: Absent; the examples do not demonstrate the use of delimiters to isolate user input from the agent's instructions. Capability inventory: The agent is granted access to the 'eval()' capability through its toolset. Sanitization: Absent; the provided code snippets do not include logic to validate or sanitize the expression strings before they are evaluated.

guidance