mechanistic-interpretability-nnsight
Pass
Audited by Gen Agent Trust Hub on Nov 3, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill references and provides instructions to install the
nnsightandtorchlibraries via standard package managers. These are legitimate tools used for deep learning research and model analysis. - [COMMAND_EXECUTION]: The documentation includes shell commands for installing dependencies (
pip install nnsight). These are standard operational instructions for the library's intended use case. - [DATA_EXFILTRATION]: The skill describes the use of the
remote=Trueparameter withinnnsight, which sends model tracing operations to thendif.usservice for remote execution on large-scale models. This is an intended feature of the library designed to enable research on models that exceed local hardware capabilities. - [SAFE]: Analysis of the skill reveals no malicious behavior. Credential handling for the NDIF API key follows safe practices by recommending environment variables or local configuration. No obfuscation, unauthorized privilege escalation, or persistence mechanisms were detected.
Audit Metadata