AI coding agents can create a new code execution risk when they treat externally influenced error data as trusted guidance and have access to command line tools, according to new research from ...