torchtitan
Pass
Audited by Gen Agent Trust Hub on Oct 2, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill instructs users to download model assets from HuggingFace and clone source code from the official pytorch GitHub organization. These are well-known and established sources in the machine learning community.- [COMMAND_EXECUTION]: Provides standard command-line instructions for training LLMs, including the use of torchrun, srun, and pip. These operations are consistent with the skill's primary purpose of distributed model training.- [CREDENTIALS_UNSAFE]: References the use of HuggingFace tokens for asset access but follows best practices by using placeholders and instructing users to use their own credentials rather than including hardcoded secrets.- [INDIRECT_PROMPT_INJECTION]: The skill defines workflows that ingest external identifiers such as repository IDs and configuration names.
- Ingestion points: Untrusted identifiers enter the training context via command-line arguments in SKILL.md.
- Boundary markers: CLI examples do not include explicit sanitization or boundary markers for these inputs.
- Capability inventory: The skill facilitates shell command execution and file system operations as seen in SKILL.md.
- Sanitization: No sanitization logic is visible in the provided instruction files.
Audit Metadata