GPUs can only access pinned memory that is intentionally mapped into their address space. Also, each context gets its own virtual address space on the GPU, isolated from other contexts.

There can still be issues, but it isn't quite as much of a free for all as the above comment sounds.

Fun fact, the Raspberry Pi's GPU can access everything. And to deal with that, the Mesa VC4 driver validates every shader to prevent reading other processes' stuff.

Isn't that because in that SOC they use unified memory where GPU and CPU memory is the same. This does not apply to most desktop computers or mobile phones...

Are there examples where this model has been abused to steal real data?

A long time ago, a bug in Firefox let a screenshot be taken of data outside the browser window:

"This issue allows attackers to capture screen shots of private or confidential information"

https://blog.mozilla.org/security/2011/06/16/webgl-graphics-...

I suppose this is more about reading another texture than the one you are supposed to use. GPU memory is flat, and there is no concept of process memory up there.

In the early days of WebGL some browsers leaked information via uninitialized GPU memory, so an attacker could potentially read texture data left behind by other processes.

Todays WebGL implementations take care to wipe new memory allocations with zeros before letting the untrusted script do anything with them though.

That's the same than starting a process and mallocing some memory; you will have the garbage of the previous process... Because you have no idea who owned that memory and what it was used to it would be hard to build something on top of that. This being said it's not a bad idea to zero stuff when you start to use them.

Which is why WebGL is only a subset of native GL ES and will never be as good.