Tinfoil hat time:
A year ago, I did some reverse engineering. I figured a pattern matching engine has to be good matching patterns, so I threw the messed-up output of Ghidra into Claude to clean it up and to give me an overview of what the fuck happens in a certain part.
I was absolutely surprised. Sure, it got some things wrong, but it knew way more than it should have been able to guess from that short bit of fucked up, optimised not-quite-C. Including guessing context that it should not be able to guess. This was a firmware with not-yet-released functions that, to my knowledge, shouldn't exist outside of this bit of binary firmware. At least not for us non-NDA mortals.
At the time, Mistral had a new-ish model as well, so I figured I'd give it a try. It worked as expected, cleaned up things, gave variables reasonable names, made educated guesses and so on, but didn't act as if it had background info. And I was using Sonnet 4.x, so the small Claude which shouldn't be that much larger.
Since then, I have the conspiracy theory that the Claude people threw a bunch of proprietary sources in the training, too. I have no idea how, maybe aggressively training on user inputs? Anyway, this does not help to allalleviate my suspicions.