I ask because I’ve been dabbling in rooting various closed devices at home. I’m finding AI fantastic at finding vulnerabilities statically (from firmware upgrade binaries), but when you then ask it to use it to get root or code execution it just goes in circles.
In the end it usually ends up there’s no exploit it can find, because it can’t actually influence the affected memory in a useful way. This is especially so if that memory is in the heap. On the stack it’s a lot more predictable (especially with OOB writes).
From experience OpenAI and Anthropic have pretty heavy safeguards in recent (public) models.
the model spins because it cant test its own guesses. what fixed it for me was emulating the actual device. qemu with the vendor rootfs, real binaries running, and then every claim becomes checkable: did the daemon actually parse my input, which uid is it under, what does its own log say. once it can check things it stops guessing, getting qemu to actually boot the thing was its own little hell though, took more tries than i want to admit.
the other thing i learned the hard way: most static findings die on reachability, not on exploit primitives. so i run them in the emulator first and kill the unreachable ones cheap, before spending any time on heap work
on heap vs stack, i think that intuition is desktop shaped. device userland usually has weak or no mitigations, heap uafs are way more tractable there than youd expect