The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructions include commands to clone repositories from LocusLab (Wanda) and IST-DASLab (SparseGPT). These are well-known academic research groups in the machine learning community, and the downloads are for standard implementations associated with the skill's purpose.
[COMMAND_EXECUTION]: The skill uses git clone and pip install to set up environment dependencies. These are standard practices for installing research software and do not involve arbitrary command execution or obfuscation.
[REMOTE_CODE_EXECUTION]: While the skill downloads external repositories, it does so from well-known academic sources for the specific purpose of model compression. No patterns of piped remote shell execution (e.g., curl | bash) were found.
[DATA_EXFILTRATION]: Network operations are limited to standard package installations and model downloads from Hugging Face (meta-llama). No sensitive file access or exfiltration to unauthorized domains was detected.
[PROMPT_INJECTION]: No instructions were found that attempt to override AI safety guidelines or extract system prompts.
[INDIRECT_PROMPT_INJECTION]: The skill processes calibration data (e.g., WikiText, C4) for activation statistics. This is a low-risk ingestion surface common in machine learning workflows, and the data is not executed or interpreted as instructions.

model-pruning