The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/run_mr_brain.py executes commands using subprocess.run to verify the local environment and trigger the synthetic data generation process. The commands are constructed using structured lists of arguments, which prevents common shell injection vulnerabilities.
[EXTERNAL_DOWNLOADS]: To function, the skill downloads the NV-Generate-CTMR repository from GitHub and model weights from Hugging Face. These resources originate from the vendor's official organizations and are necessary for the skill's operation.
[REMOTE_CODE_EXECUTION]: The skill downloads and runs Python code from the NVIDIA-Medtech/NV-Generate-CTMR repository. This code is pinned to a specific commit to ensure consistency and security, and its execution is limited to the tool's intended inference functions.
[INDIRECT_PROMPT_INJECTION]: The skill ingests user-provided JSON files to modify generation parameters, creating a potential surface for injection.
Ingestion points: Input is accepted via the model_config argument in scripts/run_mr_brain.py.
Boundary markers: The skill requires the input to be a valid JSON file.
Capability inventory: The script generates configuration files and executes subprocesses.
Sanitization: Input is strictly sanitized by filtering keys against a specific whitelist (OVERRIDE_KEYS) and performing comprehensive type and range validation on all values before execution.
[OBFUSCATION]: A pattern of using int("...") for numerical constants is present throughout the Python scripts. This appears to be a stylistic choice or a minor attempt to circumvent basic static analysis of literal values, but it does not conceal any malicious functionality.

nv-generate-mr-brain