moe-training

Pass

Audited by Gen Agent Trust Hub on Oct 29, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill instructs the user to install standard machine learning libraries like deepspeed, transformers, and accelerate from official package registries. It also provides instructions to clone the Megatron-DeepSpeed repository from Microsoft's official GitHub organization. These are documented neutrally as standard setup steps for model training.
  • [COMMAND_EXECUTION]: The documentation includes various shell commands for model training using the deepspeed CLI. These examples use typical training parameters (e.g., --num-experts, --hidden-size) and are appropriate for the skill's stated purpose of providing a technical training guide.
  • [INDIRECT_PROMPT_INJECTION]: The skill describes processes for training models on large-scale datasets, such as 'The Pile'. While processing external data is a standard part of the machine learning workflow, the skill does not include any logic that would allow untrusted data to execute commands or influence the agent's behavior in an unsafe manner. This is considered a low-risk data ingestion surface typical of developer-oriented tools.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 29, 2026, 03:22 PM