The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes metadata and artifacts from medical AI inference runs, creating a surface for indirect prompt injection where malicious instructions could be embedded in the source data.
Ingestion points: The script scripts/export_evidence_pack.py reads and processes content from manifest.json, output.json, and other evidence pack documents.
Boundary markers: The script does not utilize explicit delimiters to separate data from instructions when logging to MLflow.
Capability inventory: The skill has network and file_write permissions and uses the mlflow library to upload parameters and metrics to a remote server.
Sanitization: A robust _sanitize function is implemented to redact likely secrets, credentials, and medical identifiers (PII/PHI) such as MRNs or patient IDs from the metadata bundle before export.
[DYNAMIC_EXECUTION]: The script scripts/export_evidence_pack.py uses importlib.import_module to lazily load its primary dependencies, including mlflow, numpy, and nibabel. This is used to handle optional processing of NIfTI images and interaction with the tracking server.
[COMMAND_EXECUTION]: The skill requires the execution of a Python script through the shell to perform its core evidence processing and network export tasks.

medtech-model-evidence-export