upvote
No this is correct I think. And it's not easy to control what lives next to that buffer. But AI probably can figure out a chain of requests that this then changes the length block of another object or flips a pointer or something, so it's tedious but can be used to escalate to basically arbitrary memory corruption. That's kinda the issue with AI, this would have been extremely tedious to exploit before LLMs but now anyone with an unlocked good enough model can do it.

That's why hiding the POC still helps a bit, because any model probably will require quite a bit of wall clock time to actually build one itself. Then making that work on e.g. a running gitlab instance with other traffic... much harder yet again I'd assume since that should alter memory state without you having control over it.

Still if someone builds an exploit with even 1% success rate they just need to run it on enough instances in parallel and will eventually break through some of them probably.

reply
Right, I'm not talking down the severity of the vulnerability; if anything I think people are underrating it. But it's a constrained 1-byte overflow.
reply
Yeah that seems right on a second look
reply