SkillSecurer finds latent prompt injection in more than 17% of popular published agent skills.

A red agent generates context-compatible injections across nine threat types while recording the exact modification, and a blue agent analyzes complete skill packages and proposes patches, so a verifier scores at injection level rather than skill level (arXiv 2609.14079). With its best backend it's the only scanner in the comparison reaching 100% injection detection. Applied to popular published skills it found latent vulnerabilities in over 17% of those examined, and running those skills triggered real incidents.