distributed-llm-pretraining-torchtitan

Pass

Audited by Gen Agent Trust Hub on Feb 15, 2026

Risk Level: LOWEXTERNAL_DOWNLOADSSAFE
Full Analysis
  • [Unverifiable Dependencies & Remote Code Execution] (LOW): The skill references installing packages from the 'pytorch' GitHub organization and PyPI. These are considered trusted sources under the defined security policy, and the download finding is downgraded accordingly.
  • [Prompt Injection] (SAFE): No instructions found that attempt to bypass safety filters or override agent behavior.
  • [Data Exposure & Exfiltration] (SAFE): No unauthorized network operations or data exposure patterns were detected. Mentions of Hugging Face tokens correctly use placeholders for user-provided input.
  • [Indirect Prompt Injection] (LOW): The skill orchestrates training on external datasets. Evidence: 1. Ingestion: Dataset configurations in TOML files (e.g., file SKILL.md). 2. Boundary markers: Absent. 3. Capability inventory: Distributed execution via 'torchrun', file writing via checkpointing. 4. Sanitization: Handled by model tokenization. Severity is LOW as the content only influences internal training state without side channels for command execution.
Audit Metadata
Risk Level
LOW
Analyzed
Feb 15, 2026, 11:42 PM