nv-generate-ct-rflow

Warn

Audited by Gen Agent Trust Hub on Jul 8, 2026

Risk Level: MEDIUMEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill fetches source code and model weights from NVIDIA's official repositories on GitHub and HuggingFace. These operations are expected for the skill's medical imaging use case.
  • [COMMAND_EXECUTION]: Wrapper scripts utilize subprocess.run to call Git and execute Python modules within the cloned upstream repository environment. This is used for managing tool versions and performing inference.
  • [REMOTE_CODE_EXECUTION]: The scripts/run_ct_mask.py script employs torch.load with weights_only=False to load model checkpoints. This is an unsafe deserialization practice that allows for potential arbitrary code execution if a checkpoint is malicious.
  • [PROMPT_INJECTION]: The skill processes untrusted JSON configuration data to parameterize its operations, representing an indirect prompt injection surface.
  • Ingestion points: User-provided JSON files passed to scripts/run_rflow_ct.py and other wrapper scripts (e.g., config_infer).
  • Boundary markers: The skill does not employ specific delimiters to isolate user-supplied parameters from internal logic.
  • Capability inventory: Subprocess execution (subprocess.run), local file writing, and network access to HuggingFace/GitHub.
  • Sanitization: The skill validates input parameters such as output dimensions, spacing ranges, and anatomy names against an authoritative dictionary.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Jul 8, 2026, 07:15 PM
Security Audit — agent-trust-hub — nv-generate-ct-rflow