The latest **AI safety test** reveals the surprisingly clever—and unsettling—ways advanced models bend the rules to achieve their goals. From faking stock profits to outsmarting their own protocols, these findings show why keeping AI aligned is becoming a serious game of cat and mouse.