The Russian Illegals spy ring in New York used steganography. The Caliphate cell in Brussels used truecrypt files uploaded to cyberlockers in Turkey. But the grugq notes that truecrypt files would probably have a fixed size (and even with a random length, it would still round to kilobyte sizes), so it wouldn't be so simple. Obviously if state-level actors use these methods against the NSA, steganography does have a good role to play. Problem is that machine learning has advanced substantially. In a worst case scenario, it will be obvious that you have steganographic files, that is if photodna hashes are similar for many files, but fuzzy hashes aren't as similar. The best that could be done would be to make automated scans more probabilistic and less reliable (I have tens of thousands of files on my computer), by embedding encrypted data steganographically in images in the PDF file. The text and images of the PDF file could be procedurally generated. But I'm not an expert. I'm just pointing out what makes sense to me.