skills/davila7/claude-code-templates/mechanistic-interpretability-transformer-lens/Gen Agent Trust Hub
mechanistic-interpretability-transformer-lens
Pass
Audited by Gen Agent Trust Hub on Nov 3, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill provides instructions to install the
transformer-lenslibrary via PyPI (pip install transformer-lens) and directly from its official GitHub repository (pip install git+https://github.com/TransformerLensOrg/TransformerLens). These are standard installation methods for the core research tool described in the skill.\n- [INDIRECT_PROMPT_INJECTION]: The skill includes multiple examples that process user-provided or external text prompts to analyze transformer model internals. This constitutes an attack surface where malicious instructions in the data could potentially influence model behavior during interpretability experiments.\n - Ingestion points: Text prompts are ingested through the
model.to_tokens()method inSKILL.mdand throughout the tutorial files.\n - Boundary markers: The provided examples do not demonstrate the use of delimiters or specific safety instructions to isolate data from commands.\n
- Capability inventory: The skill utilizes the
HookedTransformerclass, which allows for reading and modifying model activations usingrun_with_hooks.\n - Sanitization: The instruction code does not implement sanitization or validation of the input strings before tokenization.
Audit Metadata