sparse-autoencoder-training
Pass
Audited by Gen Agent Trust Hub on Oct 29, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONMETADATA_POISONING
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill downloads model weights and datasets from well-known research repositories. Evidence: References to official HuggingFace models (e.g., 'gpt2-small-res-jb') and datasets (e.g., 'monology/pile-uncopyrighted'), as well as installation of the 'sae-lens' package via official registries.
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted text data for model training and activation analysis, which is a standard surface for indirect prompt injection. Ingestion points: The 'dataset_path' parameter in 'SKILL.md' and user-defined prompts in 'references/tutorials.md'. Boundary markers: None identified. Capability inventory: Includes activation steering and text generation via 'model.generate'. Sanitization: No explicit input sanitization is performed prior to tokenization.
- [METADATA_POISONING]: A discrepancy exists between the author name in the skill file and the provided author context. Evidence: The 'SKILL.md' frontmatter lists 'Orchestra Research' as the author, while the platform context identifies the author as 'xiaolai'.
Audit Metadata