The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted data, such as model responses and test inputs, which are interpolated into evaluation prompts for LLM-as-judge assessments.\n
Ingestion points: llm_judge_quality, compare_responses, and evaluate_against_reference functions in SKILL.md accept model outputs and user inputs for evaluation.\n
Boundary markers: The prompt templates for the evaluator models do not utilize explicit delimiters (like XML tags) or system instructions to ignore potential commands embedded within the data being evaluated.\n
Capability inventory: The skill interacts with external LLM APIs (Anthropic) to generate evaluation scores and reasoning.\n
Sanitization: No input sanitization or validation is performed on the data strings before they are formatted into the evaluator's prompt.\n
Context: This risk is inherent to the functionality of an evaluation framework and is considered acceptable within the skill's intended use case.

llm-evaluation