nv-generate-mr-brain
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill downloads source code from the official NVIDIA-Medtech GitHub repository and model weights from Hugging Face. These operations target trusted domains and use specific, pinned commits (
da438fe) and revisions to ensure consistency and prevent supply-chain attacks. - [COMMAND_EXECUTION]: The wrapper script (
scripts/run_mr_brain.py) usessubprocess.runto execute system commands for Git status checks and the underlying inference engine. Security is maintained by using list-based command construction (avoiding shell interpolation) and enforcing a strict allowlist of environment variables to prevent sensitive data leakage to child processes. - [INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied JSON configuration files to override model parameters. This attack surface is effectively mitigated by a dedicated validation layer (
_validate_inference_config) that enforces data types, numeric ranges, and strictly allowed keys before the data is used in the execution pipeline. - [SAFE]: The skill demonstrates multiple security best practices, such as verifying the integrity of the local upstream repository (
_upstream_identity) and calculating SHA-256 hashes of model weights to ensure they match published snapshots before execution.
Audit Metadata