The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements an 'LLM as Judge' pattern in the score_with_llm function which is susceptible to indirect prompt injection.
Ingestion points: Data is loaded from a JSON dataset file via EvaluationHarness(dataset_path) as shown in the Test Runner section of SKILL.md.
Boundary markers: The prompt template in score_with_llm lacks delimiters (such as XML tags or triple quotes) or 'ignore embedded instructions' warnings to distinguish between the grading instructions and the test data being evaluated.
Capability inventory: No high-risk capabilities (such as system command execution, network exfiltration, or privileged file access) are directly associated with the processing of the evaluation results in the provided code.
Sanitization: The expected and actual outputs are interpolated directly into the evaluation prompt using f-strings without escaping, validation, or filtering of potential adversarial instructions.

evaluation-harness