If I want to compress a text file that contains the word 'uncharacteristically' 1000 times, I can simply replace ever occurrence with the a certain short unused byte sequence and record the mapping in the header.
Similarly in an image file, if there is a block of 50 pixels that are all the same color, I can replace that block with a special byte sequence and map the coordinates and color in the header.
But if I have truly random data, there are no patterns to exploit. The shortest means of representing random data - is the data itself.
This is provable using information theory (which in a way is a branch of thermodynamics).