Also, since xor is just a CPU instruction, you won't immediately notice it in the decompiled script (if you get that far). With all the overhead that decompilers tend to produce, it's really easy to miss.
Literally the only thing that goes missing from .py to .pyc is comments.
(Not that any DRM scheme can ever work, ever, but hey. At least some try to try.)
In a pure software solution, you control the hardware, and any hiding of the key is subject to reverse engineering the software.
For all of the success they've had in protecting DirecTV, if you've got a legitimate access card feeding HDMI data out, you can make a perfect digital copy of the video stream that has no copy protection whatsoever. So ultimately the DRM offers no protection for the media content companies (at least those that don't benefit from live performances like say sports games), though it does for the pipe provider who will surely get his monthly satellite fees.
Counter-example: we once timed a release of a very minor protection update to when the main attacker typically took a holiday. We got 6 weeks out of something trivial, buying more time to work on the major release to greet him when he returned.
Popular, recently produced media has too much value to too many attackers to protect. A celebrity's self shots -- same thing. A game console by Microsoft or Sony -- same thing.
> But it's absolutely possible to make it so insanely complex and difficult that no one will ever break it
A more accurate way to put it: If you make the return on effort ratio low enough, the probability of someone breaking it goes down, and it might even go down enough for you to get away with it for a useful amount of time.
As a counter-example, I propose DirecTV or even their competitor, Dish Network (Nagravision). Hacks of these systems are worth 6 figures, pay TV is widely desired, and there hasn't been a DTV hack since 2004. None.
Putting words in my mouth there. If it's popular, the probability it will be broken goes up.
> there hasn't been a DTV hack since 2004. None.
You're preaching to the choir here. Still, that's one tidbit I didn't already know. The old mainframe Mantis language is another example.
Compilation:
1 + 2 => (+ 1 2) => push 1, push 2, add
push 1 => 1
push 2 => 2, 1
add => (+ 1 2)
Obfuscated bytecode which e.g. doesn't maintain consistent interpreter stack depths for every code path (illegal for JVM or .net CLR) would make things a little harder to analyze, but I doubt that's often the case in practice with Python.