cosmos3-post-training

Pass

Audited by Gen Agent Trust Hub on Jul 13, 2026

Risk Level: SAFE
Full Analysis
  • [SAFE]: No malicious patterns or security risks were identified during the analysis. The skill focuses on providing legitimate workflows for model fine-tuning and checkpoint management.
  • [COMMAND_EXECUTION]: The skill instructs the agent to execute local scripts and tools, such as uv, torchrun, bash, and specific Python modules (cosmos_framework.scripts.train). These actions are standard for deep learning training workflows and are executed within the user's project context.
  • [EXTERNAL_DOWNLOADS]: The skill describes downloading model weights and datasets from Hugging Face using the hf CLI tool. Hugging Face is a well-known and reputable service for hosting AI models and datasets.
  • [CREDENTIALS_UNSAFE]: The instructions mention the use of WANDB_API_KEY for Weights & Biases logging but follow best practices by advising the user to export it as an environment variable rather than hardcoding it.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 13, 2026, 10:45 PM
Security Audit — agent-trust-hub — cosmos3-post-training