The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process user-supplied data such as system prompts, model configurations, and representative inputs for the purpose of testing LLM behavior. This constitutes a surface for indirect prompt injection where untrusted data enters the agent context.
Ingestion points: Untrusted data enters via user-provided prompts, historical failures, and safety policies mentioned in prompts/llm-testing.md.
Boundary markers: There are no explicit boundary markers or instructions to ignore embedded instructions in the processed data.
Capability inventory: The skill does not perform network operations, file writes, or subprocess executions.
Sanitization: No sanitization or escaping of external content is specified.
[GENERAL_SECURITY]: The instructions in prompts/llm-testing.md explicitly promote secure practices, such as defaulting to least privilege, using masked data, mocks, and isolated environments when performing security or privacy work. The skill also enforces recording model settings and using repetition to handle stochasticity, which are best practices for reliable LLM evaluation.

llm-testing