Benchmarks & research
SkillScreen is a specialized decision model that evaluates agent skill instruction files to screen out malicious actions before tools are executed.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
Jerrold Huang and colleagues release a model and code that score SKILL.md instruction files without executing them. At 99% attack recall, their test flags 22.6% of harmless skills. On unfamiliar-source skills, the native cutoff flags 36% of harmless files at 97.3% recall. These author-run results favor catching attacks over avoiding false alarms; the project-page animation is not live inference. A clean score does not certify a skill or its external dependencies as safe.
Use the released model and code as a first screen before installing an agent skill. It reads the skill’s instruction file and assigns a risk score, rather than running the code. Test your own mix of harmless and malicious skills: the published high-detection setting also flags many harmless files.