nnsight-remote-interpretability

Pass

Audited by Gen Agent Trust Hub on Feb 17, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTION
Full Analysis
  • [EXTERNAL_DOWNLOADS] (SAFE): The skill provides instructions to install the nnsight library from PyPI (pip install nnsight). This is a legitimate research package for neural network interpretability, with source code hosted at a verified GitHub repository (ndif-team/nnsight).
  • [REMOTE_CODE_EXECUTION] (SAFE): The library supports a remote=True execution mode that offloads computation to NDIF (National Deep Inference Facility) servers. While this involves sending computation instructions to a remote endpoint (ndif.us), it is the primary stated purpose of the tool for interacting with large models and is documented in peer-reviewed research (ICLR 2025).
  • [CREDENTIALS_UNSAFE] (SAFE): Documentation includes instructions for setting an API key via environment variables (NDIF_API_KEY). Only placeholders like 'your_key' are used in examples; no sensitive credentials are leaked within the skill files.
  • [DATA_EXFILTRATION] (SAFE): No evidence of unauthorized data access or exfiltration. Network operations are limited to the intended model loading (HuggingFace) and remote inference (NDIF).
Audit Metadata
Risk Level
SAFE
Analyzed
Feb 17, 2026, 04:31 PM