The Agent Skills Directory

[DATA_EXFILTRATION]: The skill is designed to export inference metadata and artifacts to remote MLflow or Databricks servers. To mitigate the risk of exposing sensitive data, the scripts/export_evidence_pack.py script implements a _sanitize function. This function redacts keys matching a broad regex of sensitive terms (e.g., api_key, password, patient, dob, mrn), hashes prompt content using SHA256, and ensures file paths are stripped of local directory prefixes. Furthermore, the skill requires an explicit --confirm-medical-artifact-upload flag before medical images or masks are transmitted to a live server.
[DYNAMIC_EXECUTION]: The main script utilizes importlib.import_module and importlib.util.find_spec to load optional dependencies such as mlflow, numpy, and nibabel. These imports use hardcoded string literals and occur within a controlled logic flow for dependency checking and logging, which is a safe implementation of dynamic loading for known libraries.
[INDIRECT_PROMPT_INJECTION]: The skill processes inference result files that may contain data from untrusted sources. It manages this attack surface by applying the sanitization layer, treating imported content as data (not instructions), and enforcing a workflow that starts with a dry-run mode for user verification.
Ingestion points: Data is read from the path provided to the source argument in scripts/export_evidence_pack.py (e.g., manifest.json, output.json).
Boundary markers: The SKILL.md instructions require users to inspect the dry-run output, specifically params, metrics, and the artifact_plan before proceeding to live export.
Capability inventory: The skill uses file_read to ingest evidence, file_write for local MLflow stores, and network access to log data to tracking servers.
Sanitization: Implements regex-based redaction of credentials and Protected Health Information (PHI), hashes prompt text, and truncates long strings.

medtech-model-evidence-export