Excellent work by our security researcher Guanxing Wen (@hhj4ck) highlighting an important gap in AI agent security. Scanning skills or warning popups aren’t enough—without proper runtime permissions and sandboxing, a single missed review could compromise the entire host.(CertiK)