pathml

Pass

Audited by Gen Agent Trust Hub on Sep 15, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: Fetches model weights and research datasets from well-known services including Hugging Face, Warwick University, and Zenodo. It also retrieves bibliographic metadata from ArXiv for citations. All network operations are documented as requiring explicit user opt-in and disclosure of the endpoints.
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from slide manifests (CSV), spatial schemas (JSON), and whole-slide images. It implements robust sanitization via dedicated validation scripts (scripts/slide_manifest.py, scripts/validate_spatial_schema.py) that verify file paths, enforce strict schemas, and reject URLs or symlinks to prevent path traversal or context leakage.
  • Ingestion points: scripts/slide_manifest.py, scripts/validate_spatial_schema.py, and scripts/image_qc.py.
  • Boundary markers: Explicit warnings in SKILL.md regarding data de-identification and patient health information (PHI) boundaries.
  • Capability inventory: File system access, subprocess execution, and computational pathology library operations.
  • Sanitization: Strict schema validation, path traversal checks (rejecting dot segments and symlinks), and file size limits implemented in the shared scripts/_common.py utility.
  • [DYNAMIC_EXECUTION]: Documents the use of PathML's dataset loaders which may perform unsafe deserialization of .pt files (Pickle-based). The skill provides explicit warnings against loading untrusted artifacts and offers local JSON-based validation as a safer alternative.
  • [COMMAND_EXECUTION]: Instructs the agent to perform local package installation and native dependency management using standard package managers (uv, apt, brew) for local research environment setup.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 15, 2026, 01:46 AM
Security Audit — agent-trust-hub — pathml