model-pruning

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.80). The skill explicitly instructs the agent to load public calibration corpora (e.g., load_dataset("allenai/c4", ...) in references/wanda.md and load_dataset("wikitext", ...) in SKILL.md) and to use those untrusted, web-scraped text samples to collect activations that directly determine pruning masks and downstream pruning actions, so third‑party content can materially influence behavior.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.80). The skill explicitly instructs fetching and running remote code (git clone / pip install and running python main.py) from repositories such as https://github.com/locuslab/wanda and https://github.com/IST-DASLab/sparsegpt, which are required installation/runtime dependencies and will execute remote code.