moe-training

Pass

Audited by Gen Agent Trust Hub on Jun 29, 2026

Risk Level: SAFE
Full Analysis
  • [SAFE]: The skill serves as educational documentation for Mixture of Experts (MoE) machine learning architectures, providing implementation details for models like Mixtral and DeepSeek.
  • [EXTERNAL_DOWNLOADS]: The skill references standard machine learning libraries and tools, including instructions to clone the Microsoft Megatron-DeepSpeed repository and install well-known packages such as deepspeed, transformers, torch, and accelerate. These originate from trusted organizations and established registries.
  • [COMMAND_EXECUTION]: Example bash commands provided are typical for distributed training workflows (e.g., using deepspeed to launch training scripts). No suspicious commands like sudo, shell profile modifications, or obfuscated payloads were found.
  • [PROMPT_INJECTION]: The instructions focus purely on technical implementation and do not contain patterns designed to bypass agent safety filters or override system prompts.
Audit Metadata
Risk Level
SAFE
Analyzed
Jun 29, 2026, 12:55 AM
Security Audit — agent-trust-hub — moe-training