sparse-autoencoder-training
Pass
Audited by Gen Agent Trust Hub on Sep 5, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: Fetches pre-trained Sparse Autoencoder weights and configuration metadata from HuggingFace and the library's official GitHub repository (
jbloomAus/SAELens). This represents the primary intended functionality for model analysis. - [EXTERNAL_DOWNLOADS]: Downloads external datasets (e.g.,
monology/pile-uncopyrighted) from HuggingFace during the SAE training workflow. - [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted text data to analyze neural network activations, which could potentially contain malicious instructions targeting the model or agent.
- Ingestion points: Untrusted strings passed to
model.to_tokens()and external datasets specified viadataset_pathinSKILL.mdandreferences/tutorials.md. - Boundary markers: No specific delimiters or "ignore instructions" warnings are implemented for processed data.
- Capability inventory: Capabilities are limited to local model execution and file writing via
sae.save_model(); no arbitrary shell execution or network exfiltration tools are used. - Sanitization: Standard tokenization is used, but there is no specific filtering for adversarial content in the input text.
Audit Metadata