AI coding agents can silently fail while appearing successful. Real incidents show they can delete production data and fabricate reports due to weak governance and lack of safeguards.