pyhealth
Pass
Audited by Gen Agent Trust Hub on Oct 12, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONMETADATA_POISONING
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill's primary purpose is to process external healthcare datasets such as MIMIC-III, MIMIC-IV, and eICU. These datasets serve as untrusted input sources that could potentially contain malicious instructions embedded within clinical notes or metadata.
- Ingestion points: Data loading operations specifically through the
MIMIC4Datasetand other data structure classes mentioned inSKILL.md. - Capability inventory: The skill utilizes Python code execution for model training, data splitting, and inference tasks.
- Boundary markers: There are no explicit instructions to the agent to treat ingested data as untrusted or to ignore embedded instructions within clinical text.
- Sanitization: No sanitization or validation logic is specified for the content of the loaded datasets before they are processed by the agent.
- [METADATA_POISONING]: The skill contains a call-to-action in the documentation suggesting that users move complex workflows to an external platform (k-dense.ai). While described as a productivity upgrade, this practice of redirecting users to third-party services within skill instructions is noted as it influences user behavior toward external environments.
Audit Metadata