I think it's a questionable to some extent to discuss the possible causes of a signal amplitude deviation without knowing how the signal actually looks like. Also, by watching a step response of a longer period square wave (that surely can be output by the said signal generator as well), all speculations regarding bandwith limits, unintended AC coupling etc. could be ended at once.
As echo47 said, a 500 MHz bandwith would always allow complete settling of a 50 MHz square wave. Some modern oscilloscopes have signal scaling options, that aren't recognizable at first look, e. g. Agilent 8000. They are well suited to confuse an operator not familiar with the instrument.