The Agent Skills Directory

[SAFE]: No security issues were identified in the analyzed files. The skill serves its stated purpose of providing guidance for causal interventions in PyTorch models.
[EXTERNAL_DOWNLOADS]: The skill references the installation of pyvene from standard package registries and links to official documentation and repositories hosted by Stanford NLP. These are recognized as trusted or well-known sources.
[REMOTE_CODE_EXECUTION]: While the skill illustrates loading models and interventions from the HuggingFace Hub, this is a standard and intended use of the library for research reproducibility and involves well-known services.
[COMMAND_EXECUTION]: The shell commands provided in the documentation are restricted to standard library installation routines (pip install).
[DATA_EXFILTRATION]: There is no evidence of the skill accessing sensitive file paths or exfiltrating data to external domains.
[PROMPT_INJECTION]: The provided instructions are strictly focused on technical implementation and do not contain attempts to override agent constraints or safety guidelines.

mechanistic-interpretability-pyvene