Ten out of eleven popular open-source AI coding agents are vulnerable to decades-old shell injection tricks, rendering their hardcoded safety filters useless ...