Eric Pardee just wanted his $114 Amazon Fire HD 10 to run a smart-home dashboard. Amazon’s software kept force-shutting it down, and he couldn’t kill the offending packages without root. So he turned four frontier AI models into a reverse-engineering crew — and burned $266 in API bills finding out which one could actually pull it off.
What the AI models actually did
This wasn’t a chatbot writing tutorials. It was agentic hacking on a live device. Kimi K3 ($164) dug up an unpatched Mali GPU use-after-free bug (CVE-2022-38181). GLM-5.2 ($22) figured out why the first exploit attempts flopped. Claude spent five months diagnosing for free — then hit its safety guardrails and quit. GLM-5.3, on an $80 plan, finished the whole thing in a single day: it caught that the kernel was offset by 0x5C000, flipped SELinux to permissive, popped a root shell, and stripped ~100 Amazon packages.
Why it’s worth watching
An open-source model closed the job three closed flagships couldn’t. That’s why it hit HN’s front page — 567 points, 248 comments — and reignited the whole device-ownership fight.
You Might Also Like
- A Evolve Amazon Agentic ai Framework Tops mcp Atlas at 79 4 With Zero Human Tuning
- Glm 5 3 Scores 84 5 on Cybergym an Open Weight Model Just Beat gpt 5 6 sol at Hacking
- Siteline Just Dropped and it Might be the Analytics Tool the Agentic web Actually Needs
- Glm 5 Just Dropped and its the Open Source Model Nobody saw Coming
- Simon Willisons Agentic Engineering Patterns is the Practical Guide Weve Been Waiting for
Top AI Product 
Leave a comment