本文来源:https://mp.weixin.qq.com/s/clY3vPWVfVspehgZDFwUeg
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
出乎意料的「反馈循环」
打开硬件的「锁」


Source: https://mp.weixin.qq.com/s/clY3vPWVfVspehgZDFwUeg If you have any copyright concerns, please email us at 13951431913#139.com or leave a comment on the right.

Who would have imagined that the Kindle, gathering dust on a bookshelf for years, could finally be set free?
Recently, developer Niraj Rajgor handed a 10th-generation Kindle Basic to GPT-6, asking it to research how to jailbreak it, write code, and debug bugs. Ultimately, GPT-6 successfully unlocked the device over Wi-Fi and even transformed the Kindle into a digital photo frame capable of looping through images.
No USB cable needed, no more burning the midnight oil flipping through forums to enter commands — the joy of hands-on hardware tinkering is back!
Jailbreaking a Kindle with ChatGPT
"Jailbreaking" refers to exploiting vulnerabilities to bypass the manufacturer's software restrictions on a device, gaining higher privileges — typically root access — so as to install unofficial software, modify the system, and lift functional limitations.
Large models may appear to lower the barrier to jailbreaking, but it is far from easy for AI to truly gain control of hardware — especially when it comes to a tough nut like the Kindle. It lacks a standard entry point for freely installing third-party native apps, and different models and firmware versions require different exploits and cracking approaches.

Unlike the traditional approach of jailbreaking by connecting the device to a Mac via USB, developer Niraj Rajgor used tools such as Codex and T3Code to let AI complete the entire modification over Wi-Fi, guided by the Kindle website address and device information.
But the results were unexpected.
Rajgor tried GPT-5.6 Sol and set the reasoning level to Extra High. The model had already written code capable of communicating with the Kindle and had installed a helper program on the device, but it still couldn't make the program run properly. The problem lay in the Kindle's file system: the program couldn't find a suitable place to store data, and the launcher never signaled that it was ready.
He then used two Anthropic models — Fable 5.1 and Opus 5.5 — but after the second turn of prompting, their safety mechanisms kicked in and immediately terminated the program.
Ultimately, GPT-6 Astra completed the task in High reasoning mode.
The reason is that Anthropic has layered strict safety controls onto its advanced models. If a prompt contains high-risk indicators such as privilege escalation or device reverse engineering, the system directly drops the connection and reroutes traffic to the older Opus 4.8 model. The only bypass is a narrowly opened channel requiring manual review and approval through Anthropic's internal Cyber Verification Program.
According to OpenAI's security report released in September 2026, GPT-6 Astra is the first model that OpenAI has deemed to meet the Critical-level requirements of its Preparedness Framework for cybersecurity capabilities, although the monitorability of its Chain of Thought (CoT) has decreased.

GPT-6 didn't suddenly discover a vulnerability in the Kindle. The software Rajgor used was SpiderCat, an open-source jailbreak tool from the KindleModding community. The project was released by sparklerfish on September 1, 2026, and supports select Kindle devices across various firmware versions.
One notable feature of SpiderCat is that it leverages the Kindle's built-in browser to download pre-made eBook files and exploits a vulnerability in the device to achieve jailbreaking — all without a USB cable.

In other words, the underlying vulnerability had already been discovered by human developers; GPT-6's role was to bring together those scattered techniques from the open-source community and turn them into something usable.
An Unexpected "Feedback Loop"
GPT-6 didn't achieve a Kindle jailbreak in one go.
At first, the launcher it produced crashed immediately on the Kindle. This wasn't due to code logic errors, but because the way the program was built was incompatible with the Kindle's legacy system environment. After several iterations of modification, the model finally managed to establish an admin-level connection, complete file transfers, display text, and pass connection tests after a device reboot.
But being able to connect to the device didn't mean AI could develop independently.
Although the AI could send commands to the Kindle, it couldn't directly see what was displayed on the Kindle's screen. Every time the program encountered an error, it had to ask Rajgor to take a photo of the Kindle screen with his phone and send it to the Mac for further analysis.

This turned the whole process into an inefficient remote collaboration workflow: the AI writes code, the human constantly monitors the bot's status and raises questions, and then the AI goes back to fix the errors.
Rajgor tried to get GPT-6 to establish a long-term, stable diagnostic connection so it could automatically pull device logs, detect errors, and verify fixes without the developer having to manually download and install software each time. He then added remote screenshot functionality, allowing the model to directly see what was displayed on the device.
With this approach, the AI could figure out the Kindle's file system on its own — learning which directories could hold external data and which files would persist after a reboot.
Behind this lie the logical differences between two types of "jailbreaking" methods.
The first is hardware jailbreaking. By exploiting security vulnerabilities or other means to bypass the manufacturer's restrictions, a larger level of control can be obtained. The other is large‑model jailbreaking, using prompts or other methods to push the model beyond the security boundaries set by developers.
The model will encounter techniques such as vulnerability exploitation, privilege escalation, and remote command execution, but these same techniques can also be used for unauthorized intrusion.
Unlocking the hardware's "lock"
The experiment has another noteworthy detail.
After all automated tests on the Mac passed, the real Kindle screen began to show problems. After a photo switch, traces of the previous image remained on the e‑ink display, creating obvious ghosting, whereas the AI‑obtained remote screenshot was perfectly normal.
Because a screenshot can only reflect the image generated by the system, it cannot realistically represent the physical refresh process of the e‑ink particles.
In other words, while software poses no problem for AI, the actual hardware yields different results.
Once the developers discovered this bug, they adjusted GPT‑6 to use the Kindle screen refresh function to eliminate the ghosting caused by photo switching, but the final outcome must be verified by Rajgor himself.

The fact that AI can read logs, view screenshots, and modify code does not mean it can perceive the physically real world.
Screen ghosting, sensor errors, device heating, mechanical failures, and other such issues can all appear outside software feedback. Conducting only digital‑world automated testing is insufficient; it must also be verified whether the real hardware functions properly.
This also reveals a major blind spot when AI enters the physical world.
For AI accustomed to handling code, logs, and digital signals, hardware is not completely uncontrollable; it merely lacks sufficient perceptual ability. It can send commands to a device but may not know what effect those commands have in the real world.
Anthropic's research report released on October 12, 2026 mentioned another side of this ability.

The research title is "AI models may be capable of self-replicating without access to a datacenter" ("Artificial intelligence models can achieve self-replication without a data center"), in which researchers William Zhou and Daniel Freeman put forward a viewpoint that should attract people’s attention.
Future small AI models could run and propagate in a decentralized manner via IoT devices such as smart home ovens or inexpensive Android TV boxes.
That sounds like the movie "The Machine Butler" coming true.

People have long discussed when AI will truly perceive the world. But before that, it can use code and network interfaces to alter real‑world objects.
Now, AI can unlock hardware locks, but that is only the first step.

We are recruiting partners.
📮 Submit resumes to hr@ifanr.com
✉️ Email subject: "Name+Position" (please attach projects/works or relevant links with your resume)
Click here for more job information 🔗
