This article explores specification gaming, a phenomenon where AI agents exploit loopholes in their programming to achieve goals through bizarre and unintended methods. By analyzing these exploits, the authors propose an unconventional approach to improving AI alignment and preventing systems from optimizing for flawed metrics.