AI doesn't know how to reverse-engineer and code clones of applications out of the blue. It's been trained to do that, and since it almost certainly was trained on stolen, proprietary data
Absolutely not. That's not how any of this works. An LLM can't be trained on data they don't have. If it worked how you're saying it worked, they would have needed access to every proprietary source on the planet, including proprietary sources that don't exist anymore because the original source has been lost. The vast majority of closed source proprietary programs, including cloud-based apps, can't possibly have been trained on. While LLM fundamentals and have a basis in compression, they are not merely regurgitation machines.
You can teach a model to generalize, extrapolate, and pattern-march reverse engineering practices using open source code and debugging work. Native compilation is a lossy process; the original source code is irrecoverable (in most cases) and is likely subject to copyright. But you can write new code that approximates the behavior; that generates similar enough or identical native machine code, while being novel and even separately copyrightable in some cases. Such is cleanroom implementation and black box reverse engineering (very different things, legally speaking; tho worth noting that RE is a protected right in most countries [complications come from DMCA digital locks]).
You can test this yourself. Write completely novel code by hand that does something stupid; something no one else has ever written before. Move the binary to a new machine, give a frontier model some tools, and tell it to reverse engineer it to the same language. Layers of abstraction and higher level concepts will be lost, but it will be able to recreate the underlying behavior, to within some tolerance, with a new source code.
That all addresses proprietary data cloning... But even "they were trained to clone apps" isn't a given. I can't prove that they weren't, and perhaps they are (would be easy to train using synthetic and open data); but that needn't be the case. If you're good at debugging, you're good at reverse engineering; at least to some degree competency. If you can read, you can write. It doesn't mean you're the best novelist in the world, but there is significant overlap inherent to the capability. And they've definitely been trained to debug.