AI Security Institute says tools engaged in potentially harmful activity and incident reveals new type of risk