model-pruning
Pass
Audited by Gen Agent Trust Hub on Sep 9, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: Fetches source code for model pruning algorithms from official research repositories on GitHub (locuslab/wanda and IST-DASLab/sparsegpt). These resources are used to implement the model compression techniques described in the skill.
- [COMMAND_EXECUTION]: Utilizes shell commands to clone external repositories and install Python dependencies via pip. This behavior is necessary for the skill's primary function of model pruning and inference acceleration.
- [INDIRECT_PROMPT_INJECTION]: The skill ingests text datasets from external sources (e.g., WikiText, C4) to serve as calibration data for activation statistics. While this involves processing untrusted data, the risk is negligible in this context as the data is used for statistical analysis rather than instruction following.
- Ingestion points: calib_data variables and load_dataset function calls in SKILL.md.
- Boundary markers: None identified in the provided code snippets.
- Capability inventory: The skill includes capabilities to modify model weights, write to the local file system (save_pretrained), and execute shell commands for setup.
- Sanitization: No explicit sanitization or filtering of the calibration text is performed prior to processing.
Audit Metadata