I was having a think about this, and I suppose it's true that maybe there's a way to interfere with GPU code from a different process, though I gather AMD has already been focusing on that attack surface.
However, even if you succeed in that, I don't know if it would be a possible vector for privilege escalation, since GPU access to system memory is still routed through a MMU. So, it doesn't have free reign to read and write whatever it wants - just the data buffers that were mapped specifically mapped by the host, for communication with it.
Really? If I were studying some code for a chip that was illegally leaked, I would be worried that reporting any bugs I found would land me in hot water, simply for possessing a copy of copyright & trade secret IP. In fact, if the repo were still live or there were active torrents, I wouldn't even download it.
Besides, hardware source code probably has a higher learning curve, and security bugs are probably harder to spot than with software. Given that modern GPUs aren't even a good attack vector, in the first place, I doubt any but the most committed state actor would likely invest the time & resources in this route.