While everyone has been worried about it becoming too easy to hack into things, something I've been worried about for the last few years is it becoming too hard to hack into things. Right now, all of these things are being broken into to be jailbroken and such, but the companies building these hardware devices have access to the models too.
It seems like a distinctly possible and sad world that they might get so much better at keeping their devices locked down by hardening them with the support of LLMs that future devices become an absolute bitch to jailbreak, even moreso than they were before.
I don’t think it’s symmetrical. As a company, in a lot of cases, you get one shot to secure the thing. And then all owners collectively then get infinite time to jailbreak it.
Nah, we just start opening the phones, cutting off some PBC patches, soldering some wires, and intercepting/synthesizing signals live. Keeping phones locked is just a zero sum game.
I would think the more viable path (in the US) would be the legal system. i.e. Amazon pressuring frontier labs to add "safeguards" to protect their IP.
What you’re essentially saying is right now we have weak human written code that might have some vulnerabilities, but something written by a frontier LLM may not have those.
Not exactly imo, you could spend a crapton of money with white box pen testing with a bunch of frontier llms, and it might become a bunch harder to break into (it generally costs more to attack than it does to defend). Not to mention formal verification might become possible for more devices which might make it nearly impossible to find exploits.
It sounds like most of the work went into getting the slideshow to work, setting up a feedback loop so Astra could work out / blunder into a solution with the desired characteristics.
I was able to do something pretty similar w Claude Code 4 months ago w Opus 4.8 (https://github.com/brunomarinho/labs-kindle-artframe). It took a few attempts as Claude understands that jailbreaking is not legal.
I find that fascinating because I was able to get Claude to jailbreak my Kindles and add them to my tech ecosystem at home. I also was able to get it to use jailbreaking hacks on my Spotify car thing and implement some alternative ways for my raspberry pi to drive it.
Just a quick reminder, you have the legal right to "jailbreak" devices that you own. This is a recognized exemption and a right that has been fought over in court.
> 9. Computer Programs—Jailbreaking [47]
>
> Multiple organizations petitioned to renew the four exemptions for computer programs that enable electronic devices to interoperate with or to remove software applications (“jailbreaking”). These exemptions permit circumvention for the purpose of jailbreaking (1) smartphones and other portable all-purpose computing devices, (2) smart televisions, (3) voice assistant devices, and (4) routers and dedicated networking devices. No oppositions were filed against renewal. The petitions demonstrated the continuing need and justification for the exemption and that petitioners have personal knowledge and experience with regard to this exemption. Petitioners described how users of a variety products in each of these categories rely on this exemption to maintain functionality and security of older devices, to install alternative operation systems, and to customize software applications on electronic devices. Collectively, the petitions demonstrated that without this exemption, TPMs installed on the enumerated products would have an adverse effect on various noninfringing uses.
and
> 11. Computer Programs—Repairs of Devices Designed Primarily for Use by Consumers [51]
>
> The Electronic Frontier Foundation (“EFF”) petitioned to renew the exemption for computer programs that control devices designed primarily for use by consumers for diagnosis, maintenance, or repair of the device. The Office did not receive meaningful opposition to renewal.[52]
>
> The petitioners demonstrated the continuing need and justification for the exemption. For example, EFF asserted that “[m]anufacturers of these devices continue to implement technological protection measures that inhibit lawful repairs, maintenance, and diagnostics, and they show no sign of changing course.” [53]
>
> EFF has personal knowledge of this exemption, as it has been involved with the section 1201 rulemaking process since its inception and has specifically advocated for device repair exemptions.
I jailbroke my Kindle with GPT-5.5 Sol and made an external e-paper display for the long batches of text AI models generate on a regular basis.
Please note, I say so in the human note on top, but "The project code was written by GPT-5.6 Sol xhigh. It has not been verified by a human for correctness, efficiency, or speed; i.e. ymmv and use this at your personal risk."
In case you're wondering why I didn't write the code myself... I needed this as a tool to read long sequences of AI text output for work. Including code. Spending time making this tool would be counter-productive to that goal.
I do plan on writing a text rendering engine some day. Hopefully, soon.
Machine assistance also allows me to do things that would've been otherwise impossible.
For example, I'm aware of NASA's 3001 standard for human-rated spaceflight systems has very specific findings related to readability of labels. There are also other standards for checklist and document readability. Asking software to support the standards wouldn't work. Most people just aren't going to use NASA's human rating checklist to design websites and terminal outputs. They either don't know about it, or don't care enough to do so.
But machine intelligence makes it possible for me to extract NASA's learnings, turn them into requirements, and then have the machine custom build software that embeds it into my day-to-day life. I'm very grateful for that.
The world is built to a very different tune, but machine intelligence allows me to re-fit it to me.
> The project code was written by GPT-5.6 Sol xhigh. It has not been verified by a human for correctness, efficiency, or speed; i.e. ymmv and use this at your personal risk.
In an age where everyone has access to AI, AI-written code has zero value. You should share the prompt or ask your agent to write a document other people’s agents can use to reimplement the code from scratch and/or customize to their specific needs.
come on man, it has significantly reduced value, but your comment assumes that if any two people input the same prompt (or sequence of prompts) they would get the same working answer, where its very much possible he had some sequence of guiding, and the thing he is sharing is a working product. which has more value than a prompt that may produce a working product some of the time...
Meanwhile Opus 5.5 didn't even answer my question about reading user memory in a linux die notifier, a question which had nothing to do with security, just low-level system programming..
Oh wait, the author said:
> The attempts with Fable 5.1 and Opus 5.5 were both blocked by Anthropic's cybersecurity safeguards after two prompts.
Whenever I start checking a dump, the conversation with Sol won't last five rounds before it flags the work as cybersecurity-related. No matter what I do, the conversation gets cut off in the middle.
Astra is even worse. It starts saying nonsense and even refuses to run a test that we know will cause an AV (as a repro for a nullptr case).
Whenever that happens, I just switch to Opus to continue my work.
While everyone has been worried about it becoming too easy to hack into things, something I've been worried about for the last few years is it becoming too hard to hack into things. Right now, all of these things are being broken into to be jailbroken and such, but the companies building these hardware devices have access to the models too.
It seems like a distinctly possible and sad world that they might get so much better at keeping their devices locked down by hardening them with the support of LLMs that future devices become an absolute bitch to jailbreak, even moreso than they were before.
I don’t think it’s symmetrical. As a company, in a lot of cases, you get one shot to secure the thing. And then all owners collectively then get infinite time to jailbreak it.
Nah, we just start opening the phones, cutting off some PBC patches, soldering some wires, and intercepting/synthesizing signals live. Keeping phones locked is just a zero sum game.
Not to tempt fate here, but this is why eFuses were adopted at the SOC-level.
Is it just entirely unprofitable to sell hardware devices unsubsidized without locking it all into restrictive ecosystems & services?
I would think the more viable path (in the US) would be the legal system. i.e. Amazon pressuring frontier labs to add "safeguards" to protect their IP.
I'm all for allowing tinkering on a Kindle, but my pacemaker is off limits
What you’re essentially saying is right now we have weak human written code that might have some vulnerabilities, but something written by a frontier LLM may not have those.
Not exactly imo, you could spend a crapton of money with white box pen testing with a bunch of frontier llms, and it might become a bunch harder to break into (it generally costs more to attack than it does to defend). Not to mention formal verification might become possible for more devices which might make it nearly impossible to find exploits.
[delayed]
The write up says it used an existing jailbreak.
It sounds like most of the work went into getting the slideshow to work, setting up a feedback loop so Astra could work out / blunder into a solution with the desired characteristics.
I was able to do something pretty similar w Claude Code 4 months ago w Opus 4.8 (https://github.com/brunomarinho/labs-kindle-artframe). It took a few attempts as Claude understands that jailbreaking is not legal.
I find that fascinating because I was able to get Claude to jailbreak my Kindles and add them to my tech ecosystem at home. I also was able to get it to use jailbreaking hacks on my Spotify car thing and implement some alternative ways for my raspberry pi to drive it.
In the US, jailbreaking is generally legal.
Claude is just a lil bitch.
Yes! I had it refuse on a similar claim of legality and when challenged to say what law forbids it it proceeded. Lil bitch.
You just have to get it in the mood.
One more reason why regulation/market capture shouldn't be happening and leave the world reliant on one or two actors
This is what I used to jailbreak mine: https://kindlemodding.org/jailbreak-wizard.html
It will be curious how anti regulation all these execs are when they start losing money.
Just a quick reminder, you have the legal right to "jailbreak" devices that you own. This is a recognized exemption and a right that has been fought over in court.
It is your device. It is your money. You own it.
From the Federal Register, https://www.federalregister.gov/documents/2024/10/28/2024-24...
and I jailbroke my Kindle with GPT-5.5 Sol and made an external e-paper display for the long batches of text AI models generate on a regular basis.Here's the project! I MIT licensed it. https://github.com/areoformed/Re-Kindled
Please note, I say so in the human note on top, but "The project code was written by GPT-5.6 Sol xhigh. It has not been verified by a human for correctness, efficiency, or speed; i.e. ymmv and use this at your personal risk."
In case you're wondering why I didn't write the code myself... I needed this as a tool to read long sequences of AI text output for work. Including code. Spending time making this tool would be counter-productive to that goal.
I do plan on writing a text rendering engine some day. Hopefully, soon.
Machine assistance also allows me to do things that would've been otherwise impossible.
For example, I'm aware of NASA's 3001 standard for human-rated spaceflight systems has very specific findings related to readability of labels. There are also other standards for checklist and document readability. Asking software to support the standards wouldn't work. Most people just aren't going to use NASA's human rating checklist to design websites and terminal outputs. They either don't know about it, or don't care enough to do so.
But machine intelligence makes it possible for me to extract NASA's learnings, turn them into requirements, and then have the machine custom build software that embeds it into my day-to-day life. I'm very grateful for that.
The world is built to a very different tune, but machine intelligence allows me to re-fit it to me.
> The project code was written by GPT-5.6 Sol xhigh. It has not been verified by a human for correctness, efficiency, or speed; i.e. ymmv and use this at your personal risk.
In an age where everyone has access to AI, AI-written code has zero value. You should share the prompt or ask your agent to write a document other people’s agents can use to reimplement the code from scratch and/or customize to their specific needs.
come on man, it has significantly reduced value, but your comment assumes that if any two people input the same prompt (or sequence of prompts) they would get the same working answer, where its very much possible he had some sequence of guiding, and the thing he is sharing is a working product. which has more value than a prompt that may produce a working product some of the time...
Meanwhile Opus 5.5 didn't even answer my question about reading user memory in a linux die notifier, a question which had nothing to do with security, just low-level system programming..
Oh wait, the author said:
> The attempts with Fable 5.1 and Opus 5.5 were both blocked by Anthropic's cybersecurity safeguards after two prompts.
How unsurprising :)
I have the opposite experience.
Whenever I start checking a dump, the conversation with Sol won't last five rounds before it flags the work as cybersecurity-related. No matter what I do, the conversation gets cut off in the middle.
Astra is even worse. It starts saying nonsense and even refuses to run a test that we know will cause an AV (as a repro for a nullptr case).
Whenever that happens, I just switch to Opus to continue my work.
I only keep an OpenAI sub for Daybreak Blue at this point. It's pretty easy to get access to, but requires a Yubikey.
This is why I’ve gone all in on codex. I got tired of being patronized and coddled by Anthropic.
If only the hadn't cut my usage limits by 50%.
Sol 6.1 is ok