If you consider that the quality of randomness can only be assessed in proportion to the size of the data sample, you'll realise that it's not really that hard to make random numbers. Depending of course on how many you want. You have to change mindset from today. One time pads were not (and should not be) used to encrypt a 6TB hard drive's worth of porn.
I think in the early days, WW1ish time, they were made by ladies randomly typing at typewriters. If a pad's page only contains 250 numbers, that can be fairly easily be bashed out on a manual keyboard. If different individuals produce different pages, that should even out unconscious patterns. I've read this somewhere but can't find a reference. But if you try it yourself, you'll see it's easy. And how can you disprove the randomness of 250 characters that were typed randomly on purpose? You can't as historical archives show.
Later they had electro-mechanical randomising devices using a rotating cam timed with relaxation oscillators. This relies on contact bounce which is fairly random, and produced numbers at 35 baud. There's also a funky rotating disc device that I only have on my machine, so the following is an extract:-
Later during the SIGSALY days, vacuum tube diodes were used to create electrical noise which was then digitised. These (KS-3 and KW-7) devices ran at up to 100 kbaud. Here's another extract from A History of U.S. Communications Security (Volumes I and II); the David G. Boak Lectures, National Security Agency (NSA), 1973:-
And whist the British Post Office is not the NSA, there's ERNIE which is a Premium Bond picker outer machine device thing. He's moved on now to solid state diodes, but ERNIE 1 (1956) did it with these valves:-
V1 is the noisy valve. The KS-3 and KW-7 will have used a very similar circuit. Interestingly, we still use the same circuit today. We just swap the vacuum tubes for transistors.
I'm not aware of any security breaches of a OPT other than the breaking of the GEE German Diplomatic code during WW2. All of the documented instances like (Venona) relied upon duplicated key material, not poor entropy from the typist. The proof of this is that no OTPs prior to the pad duplication mistake in 1942 have been decrypted.