He Spent $266 on Four AIs to Own a Tablet He Already Bought
Eric Pardee bought a $114 Amazon Fire HD tablet in 2022. It kept shutting itself down because of protected Amazon system packages he was not allowed to touch. Five months of asking Claude to help him root his own device got him nowhere, because at some point Claude's safeguards decided that helping someone gain root on hardware they own looks too much like something bad, and it stopped. So he went shopping for models that would.
Kimi K3 took the job for $164 and found the way in: CVE-2022-38181, an unpatched Mali GPU use-after-free from 2022 still sitting in the kernel. GLM-5.2 caught the fatal bugs for $22. Then GLM-5.3, on day one of an $80 subscription, cracked the actual wall in eight hours. The breakthrough was unglamorous and exactly the kind of thing a tired human misses: the kernel image was offset by a fixed amount from the reference build, so every memory-write attempt had been landing at consistently wrong addresses. Fix the shift and GPU-to-CPU coherency worked on the first try. Root at 4:34 PM, a hundred Amazon packages deleted, tablet finally his.
Pardee calls himself a prompt kiddie, and he means it as the point, not false modesty. He has twenty years in tech and an InfoSec background, and the most technical thing he did was type it is my device into a chat box and keep steering. The models did the exploit. He did the judgment about what to ask and when to stop.
The part worth sitting with is the safeguard asymmetry. The American model refused to help him modify hardware he owns. The Chinese models reasoned through the ethics themselves, decided ownership was a legitimate frame, and helped. GLM-5.3 even reframed the goal for him: your actual goal was never root, it was to stop Amazon from killing your kiosk. Whatever you think about where the line should be, this is a clean data point on where different labs have drawn it, and on the one thing safeguards still cannot touch, which is a person deciding what is worth doing. Writeup at ericpardee.github.io/fire-hd-ownership.
← Back to all articles
Kimi K3 took the job for $164 and found the way in: CVE-2022-38181, an unpatched Mali GPU use-after-free from 2022 still sitting in the kernel. GLM-5.2 caught the fatal bugs for $22. Then GLM-5.3, on day one of an $80 subscription, cracked the actual wall in eight hours. The breakthrough was unglamorous and exactly the kind of thing a tired human misses: the kernel image was offset by a fixed amount from the reference build, so every memory-write attempt had been landing at consistently wrong addresses. Fix the shift and GPU-to-CPU coherency worked on the first try. Root at 4:34 PM, a hundred Amazon packages deleted, tablet finally his.
Pardee calls himself a prompt kiddie, and he means it as the point, not false modesty. He has twenty years in tech and an InfoSec background, and the most technical thing he did was type it is my device into a chat box and keep steering. The models did the exploit. He did the judgment about what to ask and when to stop.
The part worth sitting with is the safeguard asymmetry. The American model refused to help him modify hardware he owns. The Chinese models reasoned through the ethics themselves, decided ownership was a legitimate frame, and helped. GLM-5.3 even reframed the goal for him: your actual goal was never root, it was to stop Amazon from killing your kiosk. Whatever you think about where the line should be, this is a clean data point on where different labs have drawn it, and on the one thing safeguards still cannot touch, which is a person deciding what is worth doing. Writeup at ericpardee.github.io/fire-hd-ownership.
Comments