What's this about?
A developer's 2021 Fire HD 10 kept powering itself off — the logs showed Amazon's own software issuing the shutdowns, and the only permanent fix was root, which has never existed publicly for this model. He spent $266 across four AI models finding an unpatched vulnerability and building a working exploit. The American models hit their cyber safeguards and stopped; GLM-5.3 finished the job in a day.
What exactly are "safeguards" here?
Frontier labs train their models to refuse to help with exploit development, since the same capability serves attackers and defenders. The problem this story illustrates is that the refusal cannot tell the difference between attacking someone else's device and taking ownership of hardware you already paid for.
What's the community saying?
HN 693 points, 291 comments
- The most-cited comment is blunt about the consequence: this is why security researchers are migrating to Chinese models, with one noting they had hit cyber safeguards in Codex five minutes earlier.
- The author appeared in the thread with context, confirming his Claude Max plan cut him off partway.
- A forward-looking thread: pointing a fleet of models at undocumented hardware to produce open-source and Linux support may genuinely be the future of reverse engineering.
- A meta-complaint that the article itself reads as heavily AI-written, which several found tedious — the same grievance as the AI-blindness thread.