The Agent Skills Directory

[COMMAND_EXECUTION]: In SKILL.md, a code example for a custom calculator tool uses the Python eval() function to process input strings. Although marked as a simplified example, using eval() on unsanitized strings is an unsafe practice that can lead to arbitrary code execution if the input is controlled by an untrusted source.\n- [PROMPT_INJECTION]: The skill documents multiple multi-agent workflow patterns that ingest untrusted user data without explicit sanitization or boundary markers, which represents a potential surface for indirect prompt injection.\n
Ingestion points: User inputs enter the agent context through state fields such as user_input and customer_query as described in SKILL.md.\n
Boundary markers: The provided prompt templates in the examples lack clear delimiters or specific instructions to the model to ignore instructions embedded within the user-provided data.\n
Capability inventory: The documented workflows possess capabilities that could be abused if compromised, including web search tools and mathematical execution.\n
Sanitization: The provided architectural examples do not include validation or sanitization logic, leaving the responsibility of securing the input entirely to the user.

langgraph