The Agent Skills Directory

[Indirect Prompt Injection] (LOW): The score_with_llm function in SKILL.md implements an 'LLM-as-a-judge' pattern that is susceptible to indirect prompt injection. If the model output being evaluated contains instructions to ignore the grading rubric, the judge LLM may be influenced.
Ingestion points: Untrusted data enters the agent context through the actual variable (output from the model being tested) and the dataset_path file read in the EvaluationHarness class.
Boundary markers: Absent. The prompt template in score_with_llm interpolates {actual} and {expected} directly into the instruction string without using XML tags, triple backticks, or other delimiters to isolate the untrusted content.
Capability inventory: The provided scripts are limited to data processing and scoring; they do not contain subprocess calls, network operations, or file-writing capabilities.
Sanitization: Absent. No escaping or validation is performed on the output being graded before it is sent to the LLM judge.

evaluation-harness