SkillCloak technique evades AI agent skill scanners

The strongest variant slipped past every tested scanner. Simple packing changes leave malicious agent skills fully functional. Researchers from HKUST demonstrated the method against static analysis tools.