The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides evaluation strategies that ingest untrusted text (model responses and context) and process them using LLM judges or metrics. This creates an attack surface for indirect prompt injection where instructions embedded in the evaluated text could attempt to override the judge's behavior or bias its scores.\n
Ingestion points: Untrusted data is ingested in calculate_groundedness, llm_judge_quality, and compare_responses functions through variables like response and context.\n
Boundary markers: The skill uses simple text labels (e.g., Response: {response}) to separate fields in prompts but lacks strong delimiters or explicit instructions for the judge to ignore embedded content.\n
Capability inventory: The skill uses the openai library to perform LLM-based evaluation, allowing for the execution of logic based on the untrusted input.\n
Sanitization: No logic for sanitizing or validating inputs is implemented to prevent or detect prompt injection within the processed text.

llm-evaluation