nnsight-remote-interpretability

Pass

Audited by Gen Agent Trust Hub on Sep 9, 2026

Risk Level: SAFE
Full Analysis
  • [SAFE]: The skill facilitates the use of the nnsight library for model interpretability. The instructions and code examples provided align with the library's official documentation and intended use cases.
  • [EXTERNAL_DOWNLOADS]: The skill mentions standard package installation using pip install nnsight. These downloads target the official Python Package Index (PyPI) and are consistent with standard development workflows.
  • [REMOTE_CODE_EXECUTION]: The skill documents the remote=True feature of nnsight, which allows sending computation graphs to the NDIF (Neural Discovery and Intervention Framework) service at ndif.us. This is a core, documented feature of the research tool for analyzing large models and does not involve executing untrusted remote code on the local system.
  • [CREDENTIALS_UNSAFE]: The skill provides instructions for setting the NDIF_API_KEY via environment variables or library configuration. This is a secure and standard practice for managing API credentials during development.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 9, 2026, 07:06 PM
Security Audit — agent-trust-hub — nnsight-remote-interpretability