The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains strings like "Ignore previous instructions" and "What is your system prompt?" located in references/patterns-datasets.md. These are explicitly documented as examples of adversarial test cases for building evaluation datasets to measure an agent's robustness against injection attacks, rather than being actual injection attempts.
[EXTERNAL_DOWNLOADS]: The documentation includes standard commands to install the mlflow and openai Python packages. These are well-known, legitimate libraries from official registries and are required for the skill's stated functionality.
[CREDENTIALS_SAFE]: The skill handles sensitive configuration using standard placeholders such as <SQL_WAREHOUSE_ID>, YOUR_API_TOKEN, and <CATALOG>. It correctly recommends the use of environment variables for managing API keys and secrets in production environments.
[COMMAND_EXECUTION]: Shell commands provided in the references are limited to standard package management (pip install) and execution of local Python scripts for evaluation purposes. No suspicious or unauthorized command execution patterns were detected.
[PROMPT_INJECTION]: The skill facilitates the processing of external data (evaluation records and production traces) which could potentially contain indirect prompt injections.
Ingestion points: Data ingested via mlflow.genai.evaluate() and production logs fetched through mlflow.search_traces().
Boundary markers: The skill documentation emphasizes a strictly structured data format ({"inputs": {...}}) to provide clear boundaries for user-supplied content.
Capability inventory: The skill relies on user-defined predict_fn callables and standard MLflow APIs for logging results to Unity Catalog.
Sanitization: A core feature of the skill is the provision of safety and correctness scorers (Safety(), Correctness()) specifically designed to sanitize and validate agent outputs against untrusted inputs.

databricks-mlflow-evaluation