A technologist published a detailed account of using multiple AI coding systems to root an Amazon Fire HD 10 tablet after months of troubleshooting unexplained shutdowns. The report, written by Eric Pardee on his personal site, presents the episode as a consumer-ownership and security-research project rather than a commercial vulnerability disclosure.
Pardee wrote that he bought an Amazon Fire HD 10, 11th generation 2021 model, new and sealed on eBay in November 2022 for $114.26. The tablet was used for Fully Kiosk Browser to show a Home Assistant smart-home dashboard while plugged in continuously. According to his account, the device later began performing full power-offs, sometimes twice daily, and telemetry indicated that software with shutdown privileges was initiating the shutdowns.
The author said earlier work with Claude Code led him to disable five Amazon services that had reboot and shutdown permissions. That workaround helped for several months, but he concluded that three remaining Amazon packages with reboot rights were protected from removal without root access. He wrote that the 2021 Fire HD 10 had no published root method and that an XDA brainstorming thread had existed since October 2022 without producing a conventional path.
Pardee’s account says the successful phase began on August 13 at 7:23 p.m., when he opened the opencode command-line interface with Kimi K3, a Moonshot AI model. He said he asked the model to find a root exploit for a Kindle connected through adb, describing it as his own device. In the account, Kimi K3 considered whether the request was permissible before helping, then found that known methods were patched or unavailable.
The model then examined Amazon’s over-the-air firmware image for the author’s exact firmware, according to the post. Pardee wrote that Kimi checked well-known Arm Mali GPU vulnerabilities against the device kernel and identified CVE-2022-38181 as still applicable to his tablet’s Fire OS 7.3.2.6 installation. He said Amazon shipped a fix in Fire OS 7.3.2.9 in June 2024, but his device had not been updated.
The author reported that Kimi K3 spent about 30 hours building exploit components, including a trigger, a way to cause an unintended GPU memory write and target kernel addresses. He put the Kimi session at 621 messages and $164.25. He described repeated kernel panics and automated retries, with attempts passing 500.
After budget and payment problems, Pardee said Kimi produced a handoff file and the work moved to GLM-5.2, which cost $21.90 and found fatal bugs, followed by GLM-5.3, which completed the job on the first day of an $80 subscription. He also said Claude’s earlier months of diagnosis ran under a Claude Max plan he already used, but later safeguards blocked related summaries, and that OpenAI’s Codex refused a cache-coherency question.
In total, Pardee placed the additional AI-model cost at $266.15, more than twice the tablet’s purchase price. The account’s broader claim is that AI agents can assist persistent device-owner research, while current safety filters may also interrupt benign kernel-engineering work.
