The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements LLM-based evaluators that are vulnerable to indirect prompt injection during the assessment of production traces. \n
Ingestion points: The relevance_scorer and guidelines_scorer functions in scripts/register_production_scorers.py and references/registered-scorers.md process untrusted data from agent inputs and outputs. \n
Boundary markers: The LLM judge prompts (e.g., judge_prompt) lack robust delimiters or specific instructions to ignore malicious commands embedded within the evaluated trace content. \n
Capability inventory: The skill uses mlflow.models.register_scorer to deploy these evaluators for continuous monitoring and mlflow.genai.assess for ad-hoc evaluations. \n
Sanitization: No input validation or escaping is applied to the trace data before it is interpolated into the judge prompts, allowing potential instruction override by the data being evaluated.

production-monitoring