sparse-autoencoder-training
Pass
Audited by Gen Agent Trust Hub on Apr 4, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADS
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill provides instructions to install the 'sae-lens' package and associated dependencies via pip. It also describes fetching pre-trained model weights and datasets from HuggingFace, which is a well-known service for machine learning assets.
- [SAFE]: The skill includes placeholders like 'hf_token' and 'username/repo-name' for operations involving external services, following standard security practices for documentation.
- [SAFE]: Indirect prompt injection surface identified: untrusted data enters the context via external datasets (e.g., 'monology/pile-uncopyrighted' in SKILL.md) and user-supplied prompts for feature analysis (references/tutorials.md). No specific boundary markers or input sanitization are implemented. However, the skill's capabilities are restricted to machine learning research operations and do not involve dangerous command execution or sensitive file access using this data.
Audit Metadata