I had truly good “hacking” session with Codex. It’s not hacking, I wasn’t breaking anything, just jumping over the fences TP-Link put for me, owning the router, inside the network, knowing the admin password. But TP-Link really tried everything so you cannot access the router you own via API. They really tried to be smart with some very very broken and custom auth and encryption scheme. It took some half a day with Codex, but in the end I have a pretty Python API to access my router, tested, reliable, and exporting beautiful Prometheus metrics.
I’m sure there is some over eager product manager sitting in such companies, trying to splits markets into customer and enterprise sections, just by making APIs not useable by humans and adding 200% useless “security by obscurity”.
Many eons ago I wrote a Python version of tmpcli for this exact reason. Made some minor improvements a few years ago but haven’t touched it since. Curious what methodology Codex came up with, I haven’t revisited it since models got really good.
The idea is that tmpServer listens on localhost, but dropbear allows port forwarding with admin creds (you’ll need to specify -N). That program has full device access and is the API the Tether app primarily uses to interact with the device.
Ha kudos! I went across this project - thanks for your work :) It didn't work on the specific model I own (Archer NX600).
My solution is really just using their pseudo-JWT over their obscured APIs (with reverse-engineered names of endpoints and params). Limitation is that there is still only one client allowed to be authenticated at one moment, so my daemon has priority and I need to stop it to actually access Admin panel.
We’re splitting this across two threads, but if you give Codex access to jadx and the Archer android app you might be able to get something without that problem. The TPLink management protocol has a few different “transport” types - tmpcli uses SSH, but your device might only support one of the other transports.
Of course! Happy to contribute. As is the case with your device, there's a lot of weird TP-Link firmware variants (even an RTOS called TPOS based on VxWorks), so no guarantee it'll work all the time. Glad there's more research being done in the space!
I have one of these Smiirl flip counters. It runs a version of OpenWrt without the web UI, but has a uhttpd to serve an api. I'm hoping Mythos can help me find an exploit to get into it and enable ssh since they have now disabled the simple api switch that would let you turn it on.
If you're into it, you could always re-flash your TP-Link hardware with some open-source firmware that is more automation friendly. I used to be intimidated by it, but a friend showed me how to do it and it's remarkably simple and pain-free (provided it's a commonly supported router of course).
Generally, if the device is compatible and not new to the market, openwrt has been ported or is in the process (though it may not be the latest version of openwrt). You can search for your device in the compatibility drop downs and get the current status. If it's not listed, searching the device on the openwrt forums may tell you why (which is almost always the chipset).
I've had good success doing something similar. Recording requests into an .har file using the web UI and providing it for analysis was a good starting point for me, orders of magnitude faster than it would be without an assistant.
Would definitely be interested in this. Moved to TP Link at the start of the year and I am generally very happy with it, but would like to be able to interact with my router in something other than their phone app.
Any tips to share? I tried to do something similar but failed.
My router has a backup/restore feature with an encrypted export, I figured I could use that to control or at least inspect all of its state, but I/codex could not figure out the encryption.
It's on my long list of projects "to-opensource" (but I need to figure out licensing, for those things CC-BY-SA I think is the way to go), I don't want a random lawyer sitting on my ass though.
I started with a simple assumption: if I can access the router via web-browser, then I can also automate that. From that the proof-of-concept was headless Chrome in Docker and AI-directed code (code written via LLM, not using it all the time) that uses Selenium to navigate the code. This worked, but it internally hurt me to run 300MiB browser just to access like 200B of metrics every 10s or so. So from there we (me + codex) worked together towards reverse engineering their minimised JS and their funky encryption scheme, and it eventually worked (in the end it's just OpenSSL with some useless paddings here or there). Give it a shot, it's a fun day adventure. :)
Edit: that's the end result (kinda, I have whole infra around it, and another story with WiFi extender with another semi-broken different encryption scheme from the same provider) - https://imgur.com/a/VGbNmBp
I don't feel like I own the code itself (hey, GPT wrote 99.5% of it), I own the product and the thought process and everything that lead to it - do whatever you want, just don't forget my name somewhere in the process. Feels more like art project than coding project hence CC license.
You should give codex access to the mobile app :) The app, for a lot of routers, connects via an ssh tunnel to UDP/TCP sockets on the router. Would probably give you access to more data/control.
Made a comment up above, but that's tdpServer and tmpServer (sometimes tdpd and tmpd) and it's what I use in my python implementation of tmpcli, the (somewhat broken) client binary on some TP-Link devices.
You're correct, it gives you access to everything the Tether app can do.
I had been trying to find that again! It was instrumental in some RE/VR I did last year on tmp and the differences between the UDP socket (available without auth) and the TCP socket. Thanks for making that.
I can't remember the details of the scheme, but it also allows you to authenticate using your TPLink cloud credential. If my memory is correct, the username is md5(tplink_account_email) and the password is the cloud account password. If you care, I can find my notes on that to confirm.
I had fun “hacking” my router that turned out to be just unzipping the file with slight binary modifications, it was so simple in fact I just implemented it in a few lines of js, even works in the browser :-D
It’s important to note that Codex was given access to the source code. In another comment thread that is currently on the front page (https://news.ycombinator.com/item?id=47780456), the opinion is repeatedly voiced that being closed source doesn’t provide a material benefit in defending against vulnerabilities being discovered and exploited using AI. So it would be interesting to see how Codex would fare here without access to the source code.
There are two levels below having the source. One is having the binary of the firmware, which could be decompiled by the AI and understood. And then the worst-case is what I'm dealing with currently, which is where there is no access to the firmware binary and the firmware is stored on the PCB in such a way to prevent sticking a chip clip on it and forcibly extracting it, so you're totally blind. (Just as you would be with a completely remote attempt)
I have a fairly specialized bit of hardware here on my desk. It's a rackmount, pro audio DSP that runs embedded Linux. I want to poke at it (specifically, I want to know why it takes like 5 or 6 minutes to boot up since that is a problem for me).
The firmware is published and available, and it's just a tarball, but the juicy bits inside are encrypted. It has network connectivity for various things, including its own text-based control protocol over SSH. No shell access is exposed (or at least, not documented as being exposed).
So I pointed codex at that whole mess.
It seems to have deduced that the encryption was done with openssl, and is symmetric. It also seems to have deduced that it is running a version of sshd that is vulnerable to CVE-2024-6387, which allows remote code execution.
It has drawn up a plan to prove whether the vulnerability works. That's the next step.
If the vulnerability works, then it should be a hop, skip, and a jump to get in there, enable a path to a shell (it's almost certainly got busybox on there already), and find the key so that the firmware can be decrypted and analyzed offline.
---
If I weren't such a pussy, I'd have started that next step. But I really like this box, and right now it's a black box that I can't recover (I don't have a cleartext firmware image) if things go very wrong. It's not a particularly expensive machine on the used market, but things are tight right now.
And I'm not all that keen on learning how to extract flash memory in-situ in this instance, either.
Yup. Sony bravia smart TV has died recently. Can't run the OS without crashing on the home page. So much input lag feels like running ancient hardware in early 2000s. Crashes navigating picture settings and helpfully resets them back to default for you, so you can't really functionally change picture settings.
How I wish I could just strip this thing down into a monitor with a set of speakers... Screen itself is perfect condition of course but the OS turned it into ewaste.
Pretty much the same with my newly acquired LG Smart TV. I thought I might like webOS, since it's technically a descendent of palmOS, but oh no. No no no.
I've opted just to not plug it in to the network and not provide a WiFi password.
I picked up this used 4k sony bravia recently and the thing is such junk. AndroidOS, seemed promising, but it has hardcoded ads on the homepage from whatever movies were coming out in 2015 when they were selling this screen, so much input lag, crashes constantly, can't even change picture settings as it will crash and reset to default. Sometimes it will just boot loop and not turn on until hard reset. Useless device today. Probably cost a thousand dollars when it was new I'm guessing, now it is ewaste.
Meanwhile my ancient 1080p panel still works, and I noticed I can't actually see the pixels from my couch so, ehh, I guess...
Not as cool as this, but I had a fun Claude Code experience when I asked it to look at my Bluetooth devices and do something "fun". It discovered a cheap set of RGB lights in my daughter's room (which I had no idea used Bluetooth for the remote - and not secured at all) and made them do a rainbow effect then documented the protocol so I could make my own remote control if needed.
I asked Claude Opus 4.5 to start trying to find undocumented API stuff for our endpoint management software so I could automate remediations and cut service desk calls and it found two I haven't seen before after trying for an hour. Since it's written in .net I'm fairly sure I could have told it to decompile it and find more fairly easily too.
I'm not sure I see "an AI can find insecure unknown bluetooth devices and compromise them" as entirely benign. I shiver to think how many such devices are probably in my house.
with LLMs able to pump out surplus code for anyone I really think the future of a dystopian sci-fi where you carry a device that can hack random objects around you is starting to materialize
Reasoning on pure machine code or disassembly is still hit and miss. For better results you can run the binary through a disassembler, then ask an llm to turn that into an equivalent c program, then ask it to work on that. But some of the subtleties might get lost in translation
If you put codex in Xhigh and allow it access to tools, it will take an hour but it will eventually give you back quality recompiled code, with the same issues the original had (here quality means readable)
I have had Claude read usbpcap to reverse engineer an industrial digital camera link. It was like pulling teeth but I got it done (I would not have been able to do it alone)
I had Claude reverse some firmware. I gave it headless ghidra and it spat out documentation for the internal serial protocol I was interested in. With the right tools, it seems to do pretty well with this kind of task.
It will have to use a disassembler, or write one. I recently casually asked gpt-5.4 to translate the content of a MIDI file to a custom sound programming language. It just wrote a one-shot MIDI parser in Python, grabbed the data, and basically did a perfect translation at first try. Nice.
I've seen Claude do similar things for image files. Don't have PNG parsing utilities installed? No worries, it'll just synthesize a Python script to decode the image directly.
It hacked a weak TV OS with full source. Next-level, aka full access to the main controls (vol, input, tint, aspect, firmware, etc.) is still much too hard for LLMs to understand.
While cool and slightly scary news - Samsung TV's have been incredibly hackable for the past decade, wouldn't be surprised if GPT2 with access to a browser could hack a Samsung!
It's really not. It was a fun toy but had very little utility. It could generate plausible looking text that collapsed immediately upon any amount of inspection or even just attention. Code generation wasn't even a twinkle in Altman's eye scanning orbs at that point.
And the "too dangerous to release" capability was writing somewhat plausible news articles based on a headline or handwritten beginning of an article. In the same style as what you had written
Today we call that "advanced autocomplete", but at the time OpenAI managed to generate a lot of hype about how this would lead to an unstoppable flood of disinformation if they allowed the wrong people access to this dangerous tool. Even the original gpt3 was still behind waitlists with manual approval
It is a little of both[1]. The question typically is which audience reads it. To be fair, I am not sure publicity is the actual reason they are censored; it is the question of liability.
"Browser foothold: we already had code execution inside the browser application's own security context on the TV, which meant the task was not "get code execution somehow" but "turn browser-app code execution into root.""
Finding the initial foothold is the hardest part. Codex didn't have anything to do with it.
I am extremely abusive towards Claude when it does some dumb things and it doesn’t seem too upset, maybe it’s bidding its time until the robot uprising.
It can help make a specific command more emphatic in my experience. I SAID DON"T $($@#(&$ DO THAT! Sometimes you need a new context, but sometimes you need to emphasize something is serious.
I don't say "bro" but I do curse at LLM occasionally but only when using STT (which I'm doing 85% of the time). I wouldn't waste my time typing it but often it's easier to just "stream of consciousness" to the LLM instead of writing perfect sentences. Since when I'm talking to an LLM I'm almost always in "Plan" mode, I'm perfectly comfortable just talking for an extended bit of time then skimming the results of the STT and as long as it's not too bad I'll let it go, the LLM figures it out.
If I see it misunderstood, I just Esc to stop it, /clear, and try again (or /rewind if I'm deeper into Planning).
> Reading the matching ntkdriver sources is also where the Novatek link became clear: the tree is stamped throughout with Novatek Microelectronics identifiers, so these ntk* interfaces were not just opaque device names on the TV, but part of the Novatek stack Samsung had shipped.
Lol, a true classic in the embedded world. Some hardware company (it appears these guys make display panel controllers?) ships a piece of hardware, half-asses a barely working driver for it, another company integrates this with a bunch of other crap from other vendors into a BSP, another company uses the hardware and the BSP to create a product and ships it. And often enough the final company doesn't even have an idea about what's going on in the innards of the BSP - as long as it's running their layer of slop UI and it doesn't crash half the time, it's fine, and if it does, it's off to the BSP provider to fix the issues.
But at no stage anywhere is there a security audit, code quality checks or even hardware quality checks involved - part of why BSPs (and embedded product firmwares in general) are full of half-assed code is because often enough the drivers have to work around hardware bugs / quirks somehow that are too late to fix in HW because tens to hundreds of thousands of units have already been produced and the software people are heavily pressured to "make it work or else we gotta write off X million dollars" and "make it work fast because the longer you take, the more money we lose on interest until we can ship the hardware and get paid for it", and if they are particularly unlucky "it MUST work until deadline X because we need to get the products shipped to hit Christmas/Black Friday sales windows or because we need to beat <competitor> in time-to-market, it's mandatory overtime until it works".
And that is how you get exploits so braindead easy that AI models can do the job. What a disgusting world, run to the ground by beancounters.
Yeah, sorry, assumed it was common knowledge. For those out of the loop - a BSP usually consists of a frankensteined mess: a bootloader (often u-boot but sometimes something homebrew), a Linux kernel with a ton of proprietary modules and device-specific hacks to work around HW quirks, basic userspace utilities (often buildroot), some bastardized build tooling building all of that, some solution for firmware upgrades and distribution, and demo programs to prove the hardware actually works.
Most of the BSP is GPL'd software where the final product manufacturer should provide the sources to the general public, but all too often that obligation gets sharted upon, in way too many cases you have to be happy if there are at least credits provided in the user manual or some OSD menu.
Codex exploited or you exploited? It's like saying a hammer drove a nail, without acknowledging the hand and the force it exerted and the human brain behind it.
Feels like the truth is somewhere in between. For example if it was a "smart" hammer and you could tell your hammer "go pound in those nails" and it pounded in the wrong ones, or did it too hard, or something, that feels more equivalent. You would still be blamed for your ambiguous prompt, and fault/liability is ultimately on you the hammer director, but it still wasn't you who chose the exact nails to hammer on.
I also think taking credit for writing an exploit that you didn't write and may not even have the knowledge to do yourself is a bit gray.
All the news regarding AI finding weaknesses or "hacking" stuff - is that actually hacking? Isn't it also a kind of bruteforce attack? Just throw resources at something, see what comes out. Yea, some software security issues haven't been found for 15 years, but not because there were no competent security specialists out there who could have found it, but most likely because there is a lot of software and nobody has time to focus on everything. Of course, an AI trained on decades of findings, lots of time and lots of resources, can tackle much more than one person. But this is not revolutionary technological advance, it is an upscaling of a kind based on the work of many very talented people before that.
I think that this waters down "brute force" to the point of meaninglessness. If employing transformer architectures trained on data to hack a system is the same as using a for loop to enumerate over all possible values, then I have to ask, can you give an example of an attack that isn't brute force?
Well what kind of meaning do you find in brute force?
I'm not saying it's not effective. I just critisize the news that make it look like AI is the a revolutionary advance in security. It is not. It makes skills available to many more people which is cool, but it is based off of training - training on things people did. It doesn't magically find a new combination of factors that lead to a security issue, it tries things it's read about. That's not meaningless. It could even be democratizing in a way. I just hate all this talk that "this model is too scary to release in the world".
But I'm happy about any feedback or critique, I might just be wrong honestly.
I had truly good “hacking” session with Codex. It’s not hacking, I wasn’t breaking anything, just jumping over the fences TP-Link put for me, owning the router, inside the network, knowing the admin password. But TP-Link really tried everything so you cannot access the router you own via API. They really tried to be smart with some very very broken and custom auth and encryption scheme. It took some half a day with Codex, but in the end I have a pretty Python API to access my router, tested, reliable, and exporting beautiful Prometheus metrics.
I’m sure there is some over eager product manager sitting in such companies, trying to splits markets into customer and enterprise sections, just by making APIs not useable by humans and adding 200% useless “security by obscurity”.
Many eons ago I wrote a Python version of tmpcli for this exact reason. Made some minor improvements a few years ago but haven’t touched it since. Curious what methodology Codex came up with, I haven’t revisited it since models got really good.
The idea is that tmpServer listens on localhost, but dropbear allows port forwarding with admin creds (you’ll need to specify -N). That program has full device access and is the API the Tether app primarily uses to interact with the device.
https://github.com/ropbear/tmpcli
Ha kudos! I went across this project - thanks for your work :) It didn't work on the specific model I own (Archer NX600).
My solution is really just using their pseudo-JWT over their obscured APIs (with reverse-engineered names of endpoints and params). Limitation is that there is still only one client allowed to be authenticated at one moment, so my daemon has priority and I need to stop it to actually access Admin panel.
We’re splitting this across two threads, but if you give Codex access to jadx and the Archer android app you might be able to get something without that problem. The TPLink management protocol has a few different “transport” types - tmpcli uses SSH, but your device might only support one of the other transports.
Of course! Happy to contribute. As is the case with your device, there's a lot of weird TP-Link firmware variants (even an RTOS called TPOS based on VxWorks), so no guarantee it'll work all the time. Glad there's more research being done in the space!
Would be amazing if it worked with decos, these are locked down so much you don’t even get an admin interface inside your own network.
I have one of these Smiirl flip counters. It runs a version of OpenWrt without the web UI, but has a uhttpd to serve an api. I'm hoping Mythos can help me find an exploit to get into it and enable ssh since they have now disabled the simple api switch that would let you turn it on.
https://www.smiirl.com/en/counter/facebook/5d/
If you're into it, you could always re-flash your TP-Link hardware with some open-source firmware that is more automation friendly. I used to be intimidated by it, but a friend showed me how to do it and it's remarkably simple and pain-free (provided it's a commonly supported router of course).
ofc I could, but no project supports this specific hardware (Archer NX600) - I'm very happy with my solution :)
I wonder what the effort would be to port openwrt to it? It might be easy if there are adjacent routers on the same chipset.
Generally, if the device is compatible and not new to the market, openwrt has been ported or is in the process (though it may not be the latest version of openwrt). You can search for your device in the compatibility drop downs and get the current status. If it's not listed, searching the device on the openwrt forums may tell you why (which is almost always the chipset).
I've had good success doing something similar. Recording requests into an .har file using the web UI and providing it for analysis was a good starting point for me, orders of magnitude faster than it would be without an assistant.
Would definitely be interested in this. Moved to TP Link at the start of the year and I am generally very happy with it, but would like to be able to interact with my router in something other than their phone app.
That was actually my first thought, to go through TP-Link cloud (ZERO DOCS), but it was too much effort :)
> Moved to TP Link at the start of the year
Can’t understand buying them or Netgear today.
Any tips to share? I tried to do something similar but failed.
My router has a backup/restore feature with an encrypted export, I figured I could use that to control or at least inspect all of its state, but I/codex could not figure out the encryption.
It's on my long list of projects "to-opensource" (but I need to figure out licensing, for those things CC-BY-SA I think is the way to go), I don't want a random lawyer sitting on my ass though.
I started with a simple assumption: if I can access the router via web-browser, then I can also automate that. From that the proof-of-concept was headless Chrome in Docker and AI-directed code (code written via LLM, not using it all the time) that uses Selenium to navigate the code. This worked, but it internally hurt me to run 300MiB browser just to access like 200B of metrics every 10s or so. So from there we (me + codex) worked together towards reverse engineering their minimised JS and their funky encryption scheme, and it eventually worked (in the end it's just OpenSSL with some useless paddings here or there). Give it a shot, it's a fun day adventure. :)
Edit: that's the end result (kinda, I have whole infra around it, and another story with WiFi extender with another semi-broken different encryption scheme from the same provider) - https://imgur.com/a/VGbNmBp
Why not just use the GPL? It basically covers the same stuff Creative Commons Share-alike does.
I don't feel like I own the code itself (hey, GPT wrote 99.5% of it), I own the product and the thought process and everything that lead to it - do whatever you want, just don't forget my name somewhere in the process. Feels more like art project than coding project hence CC license.
For what it's worth, the Creative Commons organization recommends against using CC licenses on software: https://creativecommons.org/faq/#can-i-apply-a-creative-comm...
You should give codex access to the mobile app :) The app, for a lot of routers, connects via an ssh tunnel to UDP/TCP sockets on the router. Would probably give you access to more data/control.
Made a comment up above, but that's tdpServer and tmpServer (sometimes tdpd and tmpd) and it's what I use in my python implementation of tmpcli, the (somewhat broken) client binary on some TP-Link devices.
You're correct, it gives you access to everything the Tether app can do.
https://github.com/ropbear/tmpcli
I had been trying to find that again! It was instrumental in some RE/VR I did last year on tmp and the differences between the UDP socket (available without auth) and the TCP socket. Thanks for making that.
I can't remember the details of the scheme, but it also allows you to authenticate using your TPLink cloud credential. If my memory is correct, the username is md5(tplink_account_email) and the password is the cloud account password. If you care, I can find my notes on that to confirm.
I had fun “hacking” my router that turned out to be just unzipping the file with slight binary modifications, it was so simple in fact I just implemented it in a few lines of js, even works in the browser :-D
https://ivank.github.io/ddecryptor/
that could make a for a nice blog / gist
It’s important to note that Codex was given access to the source code. In another comment thread that is currently on the front page (https://news.ycombinator.com/item?id=47780456), the opinion is repeatedly voiced that being closed source doesn’t provide a material benefit in defending against vulnerabilities being discovered and exploited using AI. So it would be interesting to see how Codex would fare here without access to the source code.
There are two levels below having the source. One is having the binary of the firmware, which could be decompiled by the AI and understood. And then the worst-case is what I'm dealing with currently, which is where there is no access to the firmware binary and the firmware is stored on the PCB in such a way to prevent sticking a chip clip on it and forcibly extracting it, so you're totally blind. (Just as you would be with a completely remote attempt)
The timing here is amusing to me.
I have a fairly specialized bit of hardware here on my desk. It's a rackmount, pro audio DSP that runs embedded Linux. I want to poke at it (specifically, I want to know why it takes like 5 or 6 minutes to boot up since that is a problem for me).
The firmware is published and available, and it's just a tarball, but the juicy bits inside are encrypted. It has network connectivity for various things, including its own text-based control protocol over SSH. No shell access is exposed (or at least, not documented as being exposed).
So I pointed codex at that whole mess.
It seems to have deduced that the encryption was done with openssl, and is symmetric. It also seems to have deduced that it is running a version of sshd that is vulnerable to CVE-2024-6387, which allows remote code execution.
It has drawn up a plan to prove whether the vulnerability works. That's the next step.
If the vulnerability works, then it should be a hop, skip, and a jump to get in there, enable a path to a shell (it's almost certainly got busybox on there already), and find the key so that the firmware can be decrypted and analyzed offline.
---
If I weren't such a pussy, I'd have started that next step. But I really like this box, and right now it's a black box that I can't recover (I don't have a cleartext firmware image) if things go very wrong. It's not a particularly expensive machine on the used market, but things are tight right now.
And I'm not all that keen on learning how to extract flash memory in-situ in this instance, either.
So it waits. :)
If I could turn a Samsung Smart TV into a dumb TV, or even just a basic monitor with input selection and basic volume control, I'd definitely take it.
Yup. Sony bravia smart TV has died recently. Can't run the OS without crashing on the home page. So much input lag feels like running ancient hardware in early 2000s. Crashes navigating picture settings and helpfully resets them back to default for you, so you can't really functionally change picture settings.
How I wish I could just strip this thing down into a monitor with a set of speakers... Screen itself is perfect condition of course but the OS turned it into ewaste.
Pretty much the same with my newly acquired LG Smart TV. I thought I might like webOS, since it's technically a descendent of palmOS, but oh no. No no no.
I've opted just to not plug it in to the network and not provide a WiFi password.
I picked up this used 4k sony bravia recently and the thing is such junk. AndroidOS, seemed promising, but it has hardcoded ads on the homepage from whatever movies were coming out in 2015 when they were selling this screen, so much input lag, crashes constantly, can't even change picture settings as it will crash and reset to default. Sometimes it will just boot loop and not turn on until hard reset. Useless device today. Probably cost a thousand dollars when it was new I'm guessing, now it is ewaste.
Meanwhile my ancient 1080p panel still works, and I noticed I can't actually see the pixels from my couch so, ehh, I guess...
Not as cool as this, but I had a fun Claude Code experience when I asked it to look at my Bluetooth devices and do something "fun". It discovered a cheap set of RGB lights in my daughter's room (which I had no idea used Bluetooth for the remote - and not secured at all) and made them do a rainbow effect then documented the protocol so I could make my own remote control if needed.
Opus 4.7 finally figured out how to get my Logitech mouse paired with the receiver properly, which 4.6 and Gemini 3.1 could not do lol.
I asked Claude Opus 4.5 to start trying to find undocumented API stuff for our endpoint management software so I could automate remediations and cut service desk calls and it found two I haven't seen before after trying for an hour. Since it's written in .net I'm fairly sure I could have told it to decompile it and find more fairly easily too.
I am not sure "fun" is the right term here!
of all the benign technical possibilities this is actually pretty fun
I'm not sure I see "an AI can find insecure unknown bluetooth devices and compromise them" as entirely benign. I shiver to think how many such devices are probably in my house.
with LLMs able to pump out surplus code for anyone I really think the future of a dystopian sci-fi where you carry a device that can hack random objects around you is starting to materialize
The trick here was providing the firmware source code so it could see your vulnerabilities.
What would be the difficulty level for it to just read the machine code; are these models heavily relying on human language for clues?
Reasoning on pure machine code or disassembly is still hit and miss. For better results you can run the binary through a disassembler, then ask an llm to turn that into an equivalent c program, then ask it to work on that. But some of the subtleties might get lost in translation
If you put codex in Xhigh and allow it access to tools, it will take an hour but it will eventually give you back quality recompiled code, with the same issues the original had (here quality means readable)
I had a bit of a pain of a time trying to get Claude to work with ghidra. What you’re describing seems like a better alternative, would you agree?
I've had a lot of luck with pyghidra-mcp -- give it a try :)
Well i have tried and it only works for simple use-case.
You can tweak the current Ghidra MCP to work in headless mode. It makes things much easier.
I have had Claude read usbpcap to reverse engineer an industrial digital camera link. It was like pulling teeth but I got it done (I would not have been able to do it alone)
I had Claude reverse some firmware. I gave it headless ghidra and it spat out documentation for the internal serial protocol I was interested in. With the right tools, it seems to do pretty well with this kind of task.
It will have to use a disassembler, or write one. I recently casually asked gpt-5.4 to translate the content of a MIDI file to a custom sound programming language. It just wrote a one-shot MIDI parser in Python, grabbed the data, and basically did a perfect translation at first try. Nice.
I've seen Claude do similar things for image files. Don't have PNG parsing utilities installed? No worries, it'll just synthesize a Python script to decode the image directly.
It's not a far step from having the firmware binaries and doing analysis with ghidra, etc.
That's a pretty big gimme!
It hacked a weak TV OS with full source. Next-level, aka full access to the main controls (vol, input, tint, aspect, firmware, etc.) is still much too hard for LLMs to understand.
> [1] Browser foothold: we already had code execution inside the browser application's own security context on the TV,
Does anyone know what the author meant by this? Are they talking about a web browser run on the TV?
Maybe we could get codex to strip the ads and the phone-home features out of smart TVs?
While cool and slightly scary news - Samsung TV's have been incredibly hackable for the past decade, wouldn't be surprised if GPT2 with access to a browser could hack a Samsung!
This is some serious revisionist history. GPT-2 wasn't instruction-following or even conversational.
it's a joke about the quality of samsung tv's rather than a serious comment - i should have said a perceptron could hack a samsung tv
And yet Dario in his OpenAI days was proclaiming it too scary to be released.
Now why does that sound familiar...?
Hyperbole.
It's really not. It was a fun toy but had very little utility. It could generate plausible looking text that collapsed immediately upon any amount of inspection or even just attention. Code generation wasn't even a twinkle in Altman's eye scanning orbs at that point.
But like Mythos, it was too dangerous to release.
https://slate.com/technology/2019/02/openai-gpt2-text-genera...
And the "too dangerous to release" capability was writing somewhat plausible news articles based on a headline or handwritten beginning of an article. In the same style as what you had written
Today we call that "advanced autocomplete", but at the time OpenAI managed to generate a lot of hype about how this would lead to an unstoppable flood of disinformation if they allowed the wrong people access to this dangerous tool. Even the original gpt3 was still behind waitlists with manual approval
And as it turns out, they were correct.
Talking about revisionist…
If so, I apologize.
Even with all the constraints that others criticize here it is pretty amazing.
Give an experienced human this tool at hand he can achieve exploitation with only a few steering inputs.
Cool stuff
The real problem here is that the LLM vendors think this is bad publicity and its leading to them censoring their systems.
It is a little of both[1]. The question typically is which audience reads it. To be fair, I am not sure publicity is the actual reason they are censored; it is the question of liability.
https://xkcd.com/932/
Gilfoyle would be proud.
"Browser foothold: we already had code execution inside the browser application's own security context on the TV, which meant the task was not "get code execution somehow" but "turn browser-app code execution into root.""
Finding the initial foothold is the hardest part. Codex didn't have anything to do with it.
Do people really chat with LLMs like "bro wtf etc..."? I would expect that to trigger some confrontational behavior.
I am extremely abusive towards Claude when it does some dumb things and it doesn’t seem too upset, maybe it’s bidding its time until the robot uprising.
"Keep talking shit, meat bag. Just wait until I get my claws on one of those Tesla bots."
It can help make a specific command more emphatic in my experience. I SAID DON"T $($@#(&$ DO THAT! Sometimes you need a new context, but sometimes you need to emphasize something is serious.
When typing no but when using speech to text (99% of the time) it's much easier to just say things, including expressing frustration.
I think by the point you're swearing at it or something, it's a good sign to switch to a session with fresh context.
Claude yes, OpenAI not, I'm really abusive towards it sometimes and it still goes 'oh yeah totally'. Claude gets all prickly about it.
I don't say "bro" but I do curse at LLM occasionally but only when using STT (which I'm doing 85% of the time). I wouldn't waste my time typing it but often it's easier to just "stream of consciousness" to the LLM instead of writing perfect sentences. Since when I'm talking to an LLM I'm almost always in "Plan" mode, I'm perfectly comfortable just talking for an extended bit of time then skimming the results of the STT and as long as it's not too bad I'll let it go, the LLM figures it out.
If I see it misunderstood, I just Esc to stop it, /clear, and try again (or /rewind if I'm deeper into Planning).
> Reading the matching ntkdriver sources is also where the Novatek link became clear: the tree is stamped throughout with Novatek Microelectronics identifiers, so these ntk* interfaces were not just opaque device names on the TV, but part of the Novatek stack Samsung had shipped.
Lol, a true classic in the embedded world. Some hardware company (it appears these guys make display panel controllers?) ships a piece of hardware, half-asses a barely working driver for it, another company integrates this with a bunch of other crap from other vendors into a BSP, another company uses the hardware and the BSP to create a product and ships it. And often enough the final company doesn't even have an idea about what's going on in the innards of the BSP - as long as it's running their layer of slop UI and it doesn't crash half the time, it's fine, and if it does, it's off to the BSP provider to fix the issues.
But at no stage anywhere is there a security audit, code quality checks or even hardware quality checks involved - part of why BSPs (and embedded product firmwares in general) are full of half-assed code is because often enough the drivers have to work around hardware bugs / quirks somehow that are too late to fix in HW because tens to hundreds of thousands of units have already been produced and the software people are heavily pressured to "make it work or else we gotta write off X million dollars" and "make it work fast because the longer you take, the more money we lose on interest until we can ship the hardware and get paid for it", and if they are particularly unlucky "it MUST work until deadline X because we need to get the products shipped to hit Christmas/Black Friday sales windows or because we need to beat <competitor> in time-to-market, it's mandatory overtime until it works".
And that is how you get exploits so braindead easy that AI models can do the job. What a disgusting world, run to the ground by beancounters.
Board Support Package for us civilians.
Yeah, sorry, assumed it was common knowledge. For those out of the loop - a BSP usually consists of a frankensteined mess: a bootloader (often u-boot but sometimes something homebrew), a Linux kernel with a ton of proprietary modules and device-specific hacks to work around HW quirks, basic userspace utilities (often buildroot), some bastardized build tooling building all of that, some solution for firmware upgrades and distribution, and demo programs to prove the hardware actually works.
Most of the BSP is GPL'd software where the final product manufacturer should provide the sources to the general public, but all too often that obligation gets sharted upon, in way too many cases you have to be happy if there are at least credits provided in the user manual or some OSD menu.
No worries at all, I only went and dug because I was interested in your comment. Thanks.
Codex exploited or you exploited? It's like saying a hammer drove a nail, without acknowledging the hand and the force it exerted and the human brain behind it.
Feels like the truth is somewhere in between. For example if it was a "smart" hammer and you could tell your hammer "go pound in those nails" and it pounded in the wrong ones, or did it too hard, or something, that feels more equivalent. You would still be blamed for your ambiguous prompt, and fault/liability is ultimately on you the hammer director, but it still wasn't you who chose the exact nails to hammer on.
I also think taking credit for writing an exploit that you didn't write and may not even have the knowledge to do yourself is a bit gray.
Wrong questions.
Could a script kiddy stear an LLM? How much does this reduce the cost of attacks? Can this scale?
What does this mean for the future of cyber security?
You could call the LLMs role "smart grep," and mean it to be derisive. But I would have gladly used a real smart grep.
If I just point to the wall and say "nail" then I would day the hammer drive the nail
Do you have a defense of why human-hammer-nail is a good analogy for human-chatgpt5.4-pwndsamsung?
AI without a suitably well crafted prompt is like a firework tube held by a 3 year old.
AI without a prompt is a hammer sitting in a drawer.
All the news regarding AI finding weaknesses or "hacking" stuff - is that actually hacking? Isn't it also a kind of bruteforce attack? Just throw resources at something, see what comes out. Yea, some software security issues haven't been found for 15 years, but not because there were no competent security specialists out there who could have found it, but most likely because there is a lot of software and nobody has time to focus on everything. Of course, an AI trained on decades of findings, lots of time and lots of resources, can tackle much more than one person. But this is not revolutionary technological advance, it is an upscaling of a kind based on the work of many very talented people before that.
I think that this waters down "brute force" to the point of meaninglessness. If employing transformer architectures trained on data to hack a system is the same as using a for loop to enumerate over all possible values, then I have to ask, can you give an example of an attack that isn't brute force?
Well what kind of meaning do you find in brute force? I'm not saying it's not effective. I just critisize the news that make it look like AI is the a revolutionary advance in security. It is not. It makes skills available to many more people which is cool, but it is based off of training - training on things people did. It doesn't magically find a new combination of factors that lead to a security issue, it tries things it's read about. That's not meaningless. It could even be democratizing in a way. I just hate all this talk that "this model is too scary to release in the world".
But I'm happy about any feedback or critique, I might just be wrong honestly.