nemo-mbridge-perf-cuda-graphs
Pass
Audited by Gen Agent Trust Hub on Jul 10, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill provides standard technical documentation and configuration guidance for NVIDIA's Megatron Bridge performance features. Analysis across all 10 threat categories confirms:
- No Prompt Injection: Instructions focus on technical configuration and do not attempt to override agent safety or identity.
- No Data Exfiltration: The skill does not access sensitive files (~/.ssh, ~/.aws) or attempt unauthorized network transfers. References to local config paths (cfg.model) are standard within the Megatron framework.
- No Obfuscation: All content is human-readable markdown and standard code blocks (Python/Bash) with no hidden Unicode, Base64 commands, or homoglyphs.
- Safe Dependencies: Uses standard tools like
uv runandpythontargeting internal repository scripts (scripts/performance/run_script.py). No untrusted external packages or remote code execution via pipes (curl|bash) are present. - Safe Command Execution: Commands provided are illustrative benchmarks for performance tuning (
run_script.py) and do not involve privilege escalation (sudo) or persistence mechanisms. - Metadata and Verification: The skill contains valid metadata, benchmark data, and a digital signature (
skill.oms.sig) from NVIDIA, ensuring authenticity and integrity.
Audit Metadata