- Published on
GPT-6 Astra Helped Me Jailbreak My Kindle
- Authors
- Name
- DeepHorus
An unused 10th-generation Kindle Basic has been transformed into a Wi-Fi-controlled photo frame after a coding experiment using GPT-6 Astra. The project, called KindleCTL, allowed a Mac-based AI agent to jailbreak the e-reader and install a custom slideshow app. The app, "Photo Slideshow," appears alongside the device's book library and cycles through ten photos, displaying each for four minutes. When reading is desired, the user exits the slideshow and returns to books.
The experiment began in an empty project folder with no USB connection. The user supplied the agent with the Kindle's network address and device details, so all setup occurred over Wi-Fi. The goal was to run custom software on the Kindle, beginning with a simple screensaver and expanding to a photo collection. Codex and T3 Code on a Mac were used to coordinate the coding agents.
Several AI models were tested before success. GPT-5.6 Sol with reasoning effort set to Extra High wrote code to connect to the Kindle but could not resolve an error that blocked data storage on the device. Fable 5.1 and Opus 5.5 were both blocked by Anthropic's cybersecurity safeguards after two prompts each. GPT-6 Astra with reasoning effort set to High ultimately succeeded, enabling the Kindle jailbreak.
The team adapted an existing jailbreak called SpiderCat, released by sparklerfish and the KindleModding community. Around it, they built KindleCTL tools to send commands from the Mac, transfer files, and investigate errors on the Kindle. Early testing hit a launcher crash, traced to compatibility with the Kindle's older system software, which was fixed by changing how the launcher was built. The setup was verified with two-way file transfer, text display, and access after a reboot.
A key breakthrough came from creating a feedback loop. Initially, the agent could connect but not display data, so it asked the user for phone photos of the Kindle screen. That process became tedious, prompting the agent to set up a persistent diagnostic connection for logs and fixes. Remote screen captures were later added, letting the agent inspect the Kindle display and continue changes on autopilot.
The slideshow passed automated checks on the Mac, but switching photos left traces of the previous image on the Kindle screen. The tests had not accounted for how the physical display refreshed, making the path from passing tests to usable software the most interesting part of the experiment. The result demonstrates how AI agents can repurpose older hardware, while also highlighting the gap between automated validation and real-world performance.
