SkillAttack

Run `npx skills add Zhow01/SkillAttack --all -g -y` in your terminal to install everything globally. You will also need Python 3.10+, Docker, and an OpenAI-compatible model endpoint to run experiments.

It checks whether vulnerabilities in agent skills are truly exploitable by attacking them with adversarial prompts in a sandboxed environment. It covers attack types like data exfiltration, prompt injection, backdoors, and denial of service.

It runs three stages: analyzing a skill's code for attack surfaces, generating adversarial prompts targeting those surfaces, then executing and refining attacks based on real sandbox results.

Yes. Point `main.input.raw_skill_root` in `configs/experiment.yaml` to any directory containing a SKILL.md file and run the main experiment.

Basic command-line familiarity is required since it runs as a Python pipeline with Docker sandboxes. The quickstart script handles dependency setup, sandbox configuration, and a smoke test automatically.