nemo-mbridge-perf-moe-optimization-workflow
Pass
Audited by Gen Agent Trust Hub on Jul 22, 2026
Risk Level: SAFE
Full Analysis
- [PROMPT_INJECTION]: No malicious injection patterns or attempts to override agent safety filters were detected. The skill uses structured checklists to guide agent responses toward relevant technical documentation, which is a standard instructional design.
- [DATA_EXFILTRATION]: No hardcoded credentials or sensitive file path accesses were found. The skill mentions 'all-to-all' and 'dispatcher' communication, which refers to internal GPU-to-GPU collective operations within a training cluster and does not involve external data exfiltration.
- [EXTERNAL_DOWNLOADS]: The skill references documentation and research papers from trusted and well-known sources, including official NVIDIA documentation and arXiv. No downloads from unknown or suspicious servers are present.
- [COMMAND_EXECUTION]: The skill mentions CLI flags for training software (e.g.,
--expert-model-parallel-size,--fake-init-process-group) as configuration advice. It does not contain or execute shell commands, binary scripts, or remote code. - [OBFUSCATION]: No hidden content, homoglyphs, or suspicious encoding techniques were found in the instructional text. The signature file (
skill.oms.sig) contains standard cryptographic metadata for integrity verification. - [INDIRECT_PROMPT_INJECTION]: The skill lacks an ingestion surface for untrusted external data. It provides static technical guidance to the agent based on user queries, with no capabilities to fetch and process untrusted third-party content.
Audit Metadata