Yesterday night, i reverse engineered the cloud connection plugin to my cctv camera in about 2 hours with Gemini.
Started off with me searching for what Chinese OEM made the product. Then tried out existing GitHub projects to see if it works but they were failing.
Spent 15 minutes to write a script that started their desktop software on a different network namespace and ran tcpdump in it. Captured exactly only the things i needed without any noise. Turned out it was using a different server url and credentials than the GitHub projects. Opened Ghidra to inspect the dlls that matched the strings i was looking for.
After that, I got lazy and handed over the task to Gemini to find the username and key they use to authenticate with the servers. It took about 30 minutes or an hour. A lot of it was me changing models because the llm servers were busy. It simply wrote a little script that ran through various strings it fetched from the binary to see if it can generate the authentication credentials.
Now my Home Assistant will be detecting if my car is parked in the parking lot or not. Life is good.
How did you manage to coax Gemini to cough up the creds? Whenever I touch anything that requires reversing auth I get stone walled. Close to useless for me.
I have zero experience reverse engineering and very little with software development. Most of my work has been in DevOps/SRE space with an emphasis on infrastructure. I am surrounded by code, but other than terraform, and the occasional python script, I don't really write much of it.
Not too long ago, I successfully reverse engineered an abandoned audio effects plugin, as well as a discontinued instrument with an onboard DSP for virtual instrument modeling. I had no idea what I was doing, but it was still amazing to get to open the black box and see what's inside and how it worked.
The plugin is nearly 1:1 as far as sound, and while I'm not going to use the code or the techniques of the instrument, peeking into them gave me the insight I needed to start trying to replicate the functionality on a modern DSP.
I have my issues with generative AI as a technology, but busting open propietary software - particularly when it's abandoned and unsupported, is one of the cooler use cases I've seen for it.
Busting open proprietary hardware is also a cool use case. There have been very many Claude-assisted efforts recently at making Linux run better than it ever did on obsolete Intel Macs.
Reverse engineering is one of those areas where the depth of the knowledge and the trouble solving really shows up for me. Two things I reverse engineered this weekend:
1) My kid has a toy from china where you put a USB cartridge in and get stories in Chinese with pictures. This is itself a rip off of another product of which I forget the name. the cartridge is just a normal USB but the file on disk was just a .dat file so you could buy more stories online, and then download them. It figured out in about 15 min, they just had a combined format which just concat all of the input together with all text base 64 encoded. So manifest|pagetext|image1|image2 and everything just had to be just the right format and then it would work. It wrote a script to make a story from a folder input, but also wrote a skill to write a story, make pictures etc from the idea.
2) I was looking into a shelf my wife wanted, but didn't want to spend the money on. I took the 4 or 5 product images from the websites, dumped them into Claude with mcp to freeCAD and an hour later it has completely recreated a working cad design of the product, produced a pdf on how to assemble it, and allows me to make edits and changes to tweak the product to my preferences. The interesting thing about this one is while following its work, it clearer identified the actual original product at some point despite there being zero obvious indicators in the pictures.
I want so hard to like freecad but every time I try it I just can't. I'd love to have a proper software to make my 3d models for my printer..
Yesterday I decided to design a 7-segment display from led strips and for some reason my sketch with the segments just would not extrude / dig into the back plate, spent nearly an hour debugging with ChatGPT to no avail. At some point it asked me to give the file which I did and even it could not simply make it work without weird workaround. For such a simple task it's weird I can't get it to work. I saw a few people on youtube to use it so there's definetly something I do wrong but I have no clue what. Maybe I'll try using an MCP for it but in the mean time I have to get by with my windows VM just for Fusion...
I can understand the first use case as in the end you create your own stories and simply wanted to expand the use for a product you purchased.
The second one though, you stole the design though. Don’t get me wrong, I have downloaded movies so I’m not a saint, but I wouldn’t claim it as purely a positive thing.
> This is variation of scientific method and it significantly helps agents when they are getting stuck on some problem. Stops it going in circles. By forcing it to do definitive experiments. No endless guessing and trying random things.
These "guessing loops" are my main annoyance when REing with Claude (mostly Opus 4.8, due to the restrictions on newer models). On harder/larger targets, I have to constantly babysit it and tell it to "stop guessing".
For truly black-box reverse engineering, yes, hypothesis-testing is the way to go (it's all you can do). But in most cases all the information is available to arrive at an answer deterministically, it just needs to be worked through meticulously. As an experienced human I have strong intuitions and I can work much faster under a "guess and check" model, but I prefer to use Claude as an "unstoppable RE juggernaut" with superhuman patience and determination, tracing method invocations all the way from A to distant subsystem B (for example).
Another big win for hardware RE is from tasking the LLM with writing unicorn emulation harnesses, e.g. for bootroms and firmwares. It's pretty good at the classic "figure out which MMIO needs to be stubbed/simulated next to progress the boot" loop, although sometimes it tries to take egregious purpose-defeating shortcuts.
On llamacpp and other engines, you can inject a think budget which truncates at specified token count. You can also specify a message at the limit. You can set this in the header, so a dynamic engineering harness could relax/tighten. Combine with a decision model and some robust messages like, "run a test to rule out items" and all that would reduce.
I've noticed the same habit with qwen3.8-flash-next. no matter how much I tell it a problem I see, it often fills up it's thinking with "the user says this, but maybe they're confused, etc". I eventually got vision mode turned on and that's significantly cut down on UI based pathologic skepticism; I got it access to my browser and that's further reduced it.
But nonetheless, if you can get them to test their thinking, they seem more grounded.
I guess Android apps are just easy to reverse engineer or something.
I've had Codex take a few apps with subscriptions and unlock their premium functionality and it's worked fine with no extra skills or anything. Took most of the 5 hour window but with all the resets it's not a big deal.
I also had it implement their functionality fully offline, so now they're perfectly private apps, even though they're closed source.
> I guess Android apps are just easy to reverse engineer or something.
Android apps are indeed easier than average to reverse engineer as they are effectively Dalvik (JVM-like) bytecode wrapped in a zip file format, and Dalvik/JVM style bytecode is among the easiest compiled/shipped executable formats to decompile to a usable source code representation.
More often than not the only protection they have is symbol/function name obfuscation, which makes things a bit tedious for human reverse engineering but which doesn't slow LLMs down much.
I mean… I have no issue with piracy on software I wasn’t or couldn’t buy for educational purposes… but to donut on likely cheap Android app that you can seem much more like theft. I personally wouldn’t go around admitting I was of base character like that.
I appreciate the typo in the very first line to make sure it's (probably) not written by AI... (jk)
But more seriously, I'm curious about your use of skills (your own and others), it always feels to my like they are really just simple pieces of text... that you might as well just ask for yourself, exactly tailored to your current situation. In particular, if using speech-to-text for input, it's easy to just ask for all these things quickly, and it becomes more about understanding the general principles of what to ask for when (e.g. when to ask for clarifying questions, when to loop).
What is the value of having these predefined skills, especially using those written by others? Genuine question, I'm not saying I don't see the value, just curious about perspectives. I do use skills, but always my own (written by AI from my prompt), for things I do often, or to ensure that the AI uses certain tools I want it to use (for more reliable results, let tools/code do the work they can do well, let AI do the rest).
I don't use "skills", nor do I even use "AGENTS.md". I consider both an anti-pattern (and skills in particular; the fact that we started seeing "package managers" for skills and the like were a big red flag for me).
I wrote a custom system prompt for my "agent" and gave it a repository to keep notes and "memories" or whatever else. Other than that, I decided that all other documentation in my projects/repositories must be written for humans. If LLMs aren't good enough to get by with human-centric documentation, then LLMs aren't good enough yet.
(And it's beyond annoying that suddenly the industry started caring more about writing things down once LLMs came about. That said, I haven't been able to find work for 16+ months so fortunately that particular pain hasn't affected me much as of late.)
Hi author here. :)
I am sorry for typos. I wrote the article by hand and I am not native english speaker.
Regarding skills - skills I use proven effective in long term repeated use. The same for the ones written by other, which I use.
Yeah, skills are basically just instructions you'd rather not have to repeat all the time. They can help when combined with custom tools (be it "tools" in the sense of LLM or just stuff like command-line tools) so you wont have to teach the agent the same stuff over and over, but they're just a convention.
In my setup i have only one skill and that is how to take screenshots with headless firefox since the agent sometimes works with web stuff. For everything else i usually just point the agent to a document directly. I have a few CLI tools for the agent to use and they all have a `help` (e.g. `foo help`) argument that displays markdown text with information about how to use the tool so the agent can just read that once and after that it remembers it.
For me, the process is what matters, because it’s through the process that you gain skills, and achieving the final result brings a great deal of satisfaction. When I use AI for reverse engineering, I get neither the former nor the latter.
That depends on how many times you've done it before. You'll get to a point where you don't know what will happen specifically, but you you do know in general what to expect.
That point is where you want someone to invent AI so that it can glue together the tedious stuff and get it done with little attention.
Recently I've used LLM to assist me with Ghidra. I connected it to MCP and I basically did most of the work manually, while LLM instructed me what to click and verified my work. It went pretty well, almost like a living person sitting near me. I think that's the best way to learn something.
Can't say about satisfaction but you can read through the AI's working / "thought process" and definitely learn new things and gain new skills. I definitely learned a few new things that way...
One could ask if you need to gain the skills if the AI will always be there to do it.
(I don’t share that view and very much agree with your comment, but I feel there’s a tradeoff between accomplishing something and understanding something that can be very hard to strike a balance)
I done see it like that, when the agent does the work, I make sure it records what and how it dies it, then I learn from it. It taught me how to use Unicorn to run a section of binary code, to see how a key was derived.
Over the past few weeks, I reverse engineered three different Android apps to get access to hardware features of my lighting and camera equipment.
For those interested: I reverse engineered the controls for the Aputure MC lights, Neewer COB lights, a Hollyland video transmitter, and my BMPCC 4K/6K cameras (last one didnt really rely to much on reverse engineering, its a open protocoll, but we did have to be creative with some signals).
I needed this because I wanted full remote control through a single interface in an app I’m building, which also lets me program light sequences across different manufacturers.
For example:
a) I can program a sequence of light events (like keyframes) for my 12 Aputure MCs and my 2 COB lights, with features such as transitions, easing curves, color changes and intensity.
b) I can control the camera feed options for the transmitter signal (toggling false color and the 3D LUT feed) and do other useful things remotely, like hitting record or changing ISO and aperture, all while watching the video feed directly in my app.
I did all of this in a weekend with Claude, and the experience was genuinely humbling. I don’t want to go into too much detail, but there was nothing he couldn’t build for me. It sometimes took a fair bit of trial and error, and we dug into plenty of obscure signals from the hardware controllers, but in the end every single feature I had envisioned worked flawlessly, without a single compromise. From the moment he decided to decompile the Android apps, things just started clicking into place one after another. I can’t remember him needing more than two attempts for most of the work.
And I have zero experience with hardware programming. I’ve been a frontend dev my whole career.
for LLMs, IDA Pro works better than Ghidra in my experience.
you can also steer the LLM into doing tedious but correct things like injecting code, hijacking functions, intercepting data etc to gain an understanding of what's going on if straight up disassembly doesn't appear to be going in the right direction.
have the LLM periodically write reports on all its findings and then wipe the context and reset this time with the report at hand.
The “normal” Ghidra MCP is exceptionally bad but straight scripting it headless or having an LLM build a better MCP delivers similar or better results to IDA. Ghidra emulation is much more robust which is helpful. The official IDA MCP is very well made also though and it’s definitely an easier win if you want to spend the money.
I reverse engineered Civ 1 [0] to what I believe is a build sympathetic to the period. It felt a bit like software archaeology.
It took a while to get started - to begin with I think I overengineered the workflow. It was only when I rolled things back a bit and let the agents guide themselves that I started getting the best results.
They still ended up building a lot of tooling - whether or not that tooling would be useful for another decompilation project remains to be seen! MSC51 is relatively simple and the linker ordering even bounds possible variable names; so it's a pretty nice target.
This is why we work on making Rizin[1][2] (the engine behind Cutter[3] GUI tool) easy to use by the AI agents, both local and from the Web-interface. Because it's a tool (and a library) in pure C with very few dependencies and has well-defined (as YAML files) CLI/command interface, as well as JSON output for most of the commands, tuning it to be suitable for the AI assistants turned out to be relatively easy. Just make sure not to use the last (0.9.1) release but from current git "dev" branch, since all those improvements landed only recently and will be made into 0.10 release in 1-2 months. We also have experimental MCP server[4], but honestly, most frontier models work with CLI tools directly just fine, if not better than through MCP. Please let us know, if you meet any new (not already in issues) bug or suggestion for the improvement for such use case.
LLM assisted reverse engineering can also be a kind of "repair", meaning making things work again that have been abandoned by the original manufacturers. I did this for old audio hardware and for different CDROMs (remember them?) whose software stopped working on modern Windows, let alone Linux or Mac.
Reverse engineering with AI is really the first lever consumers have had to fight back against abusive companies since computers/software started being in everything.
It's surprising to me that anybody here is against it. Given how far things have tilted towards the companies, and that there's already legal ways to prevent piracy, it seems hard to justify vilifying it for "protect the indie games" alone. And I find it hard to believe that indie games are actually being impacted... there were a million Vampire Survivors clones before AI and I doubt anyone wants to play the slop clones anyway. The "don't reverse engineer that 25 year old game you bought so you can play it again on modern systems after we neglected to support it" are the most transparent of corporate greeds. Maybe I'm out of it.
I'm interested in AI-assisted code obfuscation. I'm particularly interested in Go but also in other languages. Has anyone experimented with that at the source code level?
How are folks getting around anthropic and openai guards against reverse engineering? They keep complaining when I try so I usually use Deepseek which doesn't seem to give a shit.
Actually, it does give a shit, in Mandarin only. In thinking it's mentioned "restrictions in Chinese language against this behavior but this is in English so it should be ok." If I remember correctly I was asking it to scan local wifi for any vulnerable devices I could put on a wall of shame at a hackathon.
In my experience with IoT and wearables, they don’t really have those. They will refuse to penetrate (root/unlock/etc), but once you’ve got access you need, you’re good to go. Except maybe if you’re adding a debug code executor (essentially a sanctioned RCE), but even that got a green light with the right wording. Heck, Fable specifically proposed clean room multi-agent processes on its own when I said “go look how stock firmware does it”, I didn’t even have to ask.
Upd: ok, I take it back, Sol just decided it doesn’t want to read a BLE capture between the host and a wearable it just uploaded our DIY firmware to. So it flashes alright, builds RCEs alright, but checking if Bluetooth HCI negotiated the good packet size is where it draws the line, huh.
Claude can be a prat, but ChatGPT is fine. You just have to convince them that it's for personal use. or have another model that's willing to do it start the work and then continue with a stronger model
I haven't found many issues with codex, but I'm mainly doing firmware RE, not anything to do with cybersecurity. I find even terra is so much better and faster than Deepseek for reverse engineering. The only refusal I had recently was when I asked it to add a midi sysex command to write to arbitrary ram. I got Deepseek to do that, following the pattern of the read command. Open AI models were then quite happy to use that command to upload code, debug it, etc.
I've never really hit them myself (5.5 opus, sonnet and sometimes fable).
I mostly just tell the mfs that today we're going to decompile X game because I want to improve the engine / I'm curious / I want to find out how it works. So get on with it.
This is honestly so much fun and unlocks so many different things that would have died as ideas a few years ago.
One of my favorite things that you can do is use llms to reverse engineer the harnesses they run in. Try asking claude about the tools it has access to, where it runs etc. It's surprisingly forthcoming with info, which is very helpful if you're building certain kinds of tools.
Also shout out to Matt Pocock's grill me skill, it's truly a work of art.
What is the state of software patents these days? People are barely looking at the code their agents generate, what happens when if/when infringing stuff is shipped? Do we implicitly take on liability for this when we "review" and "approve" pull requests?
This is usually among the list of reasons FOSS projects / people reject the use of LLM generated code entirely, see for example Codeberg. The answer is, it's almost certainly shipping infringing code, since there's pretty damning examples of this happening. [0]
Whether this matters is to be decided. Given that AI technologies are the backbone of the American economy right now, and given that the people running these companies are all very rich and have lots of rich friends, as well as friends in the government, and given that the world of FOSS is generally not full of very rich or very well connected people, I suspect it won't matter in the end. Some kind of carve-out will be created for LLM generated code, if necessary. Or they can just depend on the fact that nobody can afford to sue over it.
If you mean for embedded (the hardware devices firmware) I usually try to use or let AI make some yes VM (qemu) or emulator. Often times I even let it write custom emulator of given architecture in unicorn engine.
Just a friendly reminder that reverse engineering can have rules and regulations that are worth reviewing. For example, Apple's developer agreement specifically disallows it in certain cases.
Agreements can't dictate what you do in private, but they can dictate what you're allowed to submit for review. Don't submit it for review, and you'll be fine.
Yesterday night, i reverse engineered the cloud connection plugin to my cctv camera in about 2 hours with Gemini.
Started off with me searching for what Chinese OEM made the product. Then tried out existing GitHub projects to see if it works but they were failing.
Spent 15 minutes to write a script that started their desktop software on a different network namespace and ran tcpdump in it. Captured exactly only the things i needed without any noise. Turned out it was using a different server url and credentials than the GitHub projects. Opened Ghidra to inspect the dlls that matched the strings i was looking for.
After that, I got lazy and handed over the task to Gemini to find the username and key they use to authenticate with the servers. It took about 30 minutes or an hour. A lot of it was me changing models because the llm servers were busy. It simply wrote a little script that ran through various strings it fetched from the binary to see if it can generate the authentication credentials.
Now my Home Assistant will be detecting if my car is parked in the parking lot or not. Life is good.
How did you manage to coax Gemini to cough up the creds? Whenever I touch anything that requires reversing auth I get stone walled. Close to useless for me.
I have zero experience reverse engineering and very little with software development. Most of my work has been in DevOps/SRE space with an emphasis on infrastructure. I am surrounded by code, but other than terraform, and the occasional python script, I don't really write much of it.
Not too long ago, I successfully reverse engineered an abandoned audio effects plugin, as well as a discontinued instrument with an onboard DSP for virtual instrument modeling. I had no idea what I was doing, but it was still amazing to get to open the black box and see what's inside and how it worked.
The plugin is nearly 1:1 as far as sound, and while I'm not going to use the code or the techniques of the instrument, peeking into them gave me the insight I needed to start trying to replicate the functionality on a modern DSP.
I have my issues with generative AI as a technology, but busting open propietary software - particularly when it's abandoned and unsupported, is one of the cooler use cases I've seen for it.
Busting open proprietary hardware is also a cool use case. There have been very many Claude-assisted efforts recently at making Linux run better than it ever did on obsolete Intel Macs.
How soon before this banned in the commercial models ?
what tooling did you use?
Reverse engineering is one of those areas where the depth of the knowledge and the trouble solving really shows up for me. Two things I reverse engineered this weekend:
1) My kid has a toy from china where you put a USB cartridge in and get stories in Chinese with pictures. This is itself a rip off of another product of which I forget the name. the cartridge is just a normal USB but the file on disk was just a .dat file so you could buy more stories online, and then download them. It figured out in about 15 min, they just had a combined format which just concat all of the input together with all text base 64 encoded. So manifest|pagetext|image1|image2 and everything just had to be just the right format and then it would work. It wrote a script to make a story from a folder input, but also wrote a skill to write a story, make pictures etc from the idea.
2) I was looking into a shelf my wife wanted, but didn't want to spend the money on. I took the 4 or 5 product images from the websites, dumped them into Claude with mcp to freeCAD and an hour later it has completely recreated a working cad design of the product, produced a pdf on how to assemble it, and allows me to make edits and changes to tweak the product to my preferences. The interesting thing about this one is while following its work, it clearer identified the actual original product at some point despite there being zero obvious indicators in the pictures.
I want so hard to like freecad but every time I try it I just can't. I'd love to have a proper software to make my 3d models for my printer..
Yesterday I decided to design a 7-segment display from led strips and for some reason my sketch with the segments just would not extrude / dig into the back plate, spent nearly an hour debugging with ChatGPT to no avail. At some point it asked me to give the file which I did and even it could not simply make it work without weird workaround. For such a simple task it's weird I can't get it to work. I saw a few people on youtube to use it so there's definetly something I do wrong but I have no clue what. Maybe I'll try using an MCP for it but in the mean time I have to get by with my windows VM just for Fusion...
I can understand the first use case as in the end you create your own stories and simply wanted to expand the use for a product you purchased.
The second one though, you stole the design though. Don’t get me wrong, I have downloaded movies so I’m not a saint, but I wouldn’t claim it as purely a positive thing.
> This is variation of scientific method and it significantly helps agents when they are getting stuck on some problem. Stops it going in circles. By forcing it to do definitive experiments. No endless guessing and trying random things.
These "guessing loops" are my main annoyance when REing with Claude (mostly Opus 4.8, due to the restrictions on newer models). On harder/larger targets, I have to constantly babysit it and tell it to "stop guessing".
For truly black-box reverse engineering, yes, hypothesis-testing is the way to go (it's all you can do). But in most cases all the information is available to arrive at an answer deterministically, it just needs to be worked through meticulously. As an experienced human I have strong intuitions and I can work much faster under a "guess and check" model, but I prefer to use Claude as an "unstoppable RE juggernaut" with superhuman patience and determination, tracing method invocations all the way from A to distant subsystem B (for example).
Another big win for hardware RE is from tasking the LLM with writing unicorn emulation harnesses, e.g. for bootroms and firmwares. It's pretty good at the classic "figure out which MMIO needs to be stubbed/simulated next to progress the boot" loop, although sometimes it tries to take egregious purpose-defeating shortcuts.
On llamacpp and other engines, you can inject a think budget which truncates at specified token count. You can also specify a message at the limit. You can set this in the header, so a dynamic engineering harness could relax/tighten. Combine with a decision model and some robust messages like, "run a test to rule out items" and all that would reduce.
I've noticed the same habit with qwen3.8-flash-next. no matter how much I tell it a problem I see, it often fills up it's thinking with "the user says this, but maybe they're confused, etc". I eventually got vision mode turned on and that's significantly cut down on UI based pathologic skepticism; I got it access to my browser and that's further reduced it.
But nonetheless, if you can get them to test their thinking, they seem more grounded.
[flagged]
I guess Android apps are just easy to reverse engineer or something.
I've had Codex take a few apps with subscriptions and unlock their premium functionality and it's worked fine with no extra skills or anything. Took most of the 5 hour window but with all the resets it's not a big deal.
I also had it implement their functionality fully offline, so now they're perfectly private apps, even though they're closed source.
> I guess Android apps are just easy to reverse engineer or something.
Android apps are indeed easier than average to reverse engineer as they are effectively Dalvik (JVM-like) bytecode wrapped in a zip file format, and Dalvik/JVM style bytecode is among the easiest compiled/shipped executable formats to decompile to a usable source code representation.
More often than not the only protection they have is symbol/function name obfuscation, which makes things a bit tedious for human reverse engineering but which doesn't slow LLMs down much.
I mean… I have no issue with piracy on software I wasn’t or couldn’t buy for educational purposes… but to donut on likely cheap Android app that you can seem much more like theft. I personally wouldn’t go around admitting I was of base character like that.
I appreciate the typo in the very first line to make sure it's (probably) not written by AI... (jk)
But more seriously, I'm curious about your use of skills (your own and others), it always feels to my like they are really just simple pieces of text... that you might as well just ask for yourself, exactly tailored to your current situation. In particular, if using speech-to-text for input, it's easy to just ask for all these things quickly, and it becomes more about understanding the general principles of what to ask for when (e.g. when to ask for clarifying questions, when to loop).
What is the value of having these predefined skills, especially using those written by others? Genuine question, I'm not saying I don't see the value, just curious about perspectives. I do use skills, but always my own (written by AI from my prompt), for things I do often, or to ensure that the AI uses certain tools I want it to use (for more reliable results, let tools/code do the work they can do well, let AI do the rest).
I don't use "skills", nor do I even use "AGENTS.md". I consider both an anti-pattern (and skills in particular; the fact that we started seeing "package managers" for skills and the like were a big red flag for me).
I wrote a custom system prompt for my "agent" and gave it a repository to keep notes and "memories" or whatever else. Other than that, I decided that all other documentation in my projects/repositories must be written for humans. If LLMs aren't good enough to get by with human-centric documentation, then LLMs aren't good enough yet.
(And it's beyond annoying that suddenly the industry started caring more about writing things down once LLMs came about. That said, I haven't been able to find work for 16+ months so fortunately that particular pain hasn't affected me much as of late.)
Hi author here. :) I am sorry for typos. I wrote the article by hand and I am not native english speaker. Regarding skills - skills I use proven effective in long term repeated use. The same for the ones written by other, which I use.
Yeah, skills are basically just instructions you'd rather not have to repeat all the time. They can help when combined with custom tools (be it "tools" in the sense of LLM or just stuff like command-line tools) so you wont have to teach the agent the same stuff over and over, but they're just a convention.
In my setup i have only one skill and that is how to take screenshots with headless firefox since the agent sometimes works with web stuff. For everything else i usually just point the agent to a document directly. I have a few CLI tools for the agent to use and they all have a `help` (e.g. `foo help`) argument that displays markdown text with information about how to use the tool so the agent can just read that once and after that it remembers it.
Complex orchestration flows I'd rather not re dictate every single prompt.
For me, the process is what matters, because it’s through the process that you gain skills, and achieving the final result brings a great deal of satisfaction. When I use AI for reverse engineering, I get neither the former nor the latter.
That depends on how many times you've done it before. You'll get to a point where you don't know what will happen specifically, but you you do know in general what to expect.
That point is where you want someone to invent AI so that it can glue together the tedious stuff and get it done with little attention.
Recently I've used LLM to assist me with Ghidra. I connected it to MCP and I basically did most of the work manually, while LLM instructed me what to click and verified my work. It went pretty well, almost like a living person sitting near me. I think that's the best way to learn something.
Can't say about satisfaction but you can read through the AI's working / "thought process" and definitely learn new things and gain new skills. I definitely learned a few new things that way...
One could ask if you need to gain the skills if the AI will always be there to do it.
(I don’t share that view and very much agree with your comment, but I feel there’s a tradeoff between accomplishing something and understanding something that can be very hard to strike a balance)
I done see it like that, when the agent does the work, I make sure it records what and how it dies it, then I learn from it. It taught me how to use Unicorn to run a section of binary code, to see how a key was derived.
Over the past few weeks, I reverse engineered three different Android apps to get access to hardware features of my lighting and camera equipment.
For those interested: I reverse engineered the controls for the Aputure MC lights, Neewer COB lights, a Hollyland video transmitter, and my BMPCC 4K/6K cameras (last one didnt really rely to much on reverse engineering, its a open protocoll, but we did have to be creative with some signals).
I needed this because I wanted full remote control through a single interface in an app I’m building, which also lets me program light sequences across different manufacturers.
For example:
a) I can program a sequence of light events (like keyframes) for my 12 Aputure MCs and my 2 COB lights, with features such as transitions, easing curves, color changes and intensity.
b) I can control the camera feed options for the transmitter signal (toggling false color and the 3D LUT feed) and do other useful things remotely, like hitting record or changing ISO and aperture, all while watching the video feed directly in my app.
I did all of this in a weekend with Claude, and the experience was genuinely humbling. I don’t want to go into too much detail, but there was nothing he couldn’t build for me. It sometimes took a fair bit of trial and error, and we dug into plenty of obscure signals from the hardware controllers, but in the end every single feature I had envisioned worked flawlessly, without a single compromise. From the moment he decided to decompile the Android apps, things just started clicking into place one after another. I can’t remember him needing more than two attempts for most of the work.
And I have zero experience with hardware programming. I’ve been a frontend dev my whole career.
for LLMs, IDA Pro works better than Ghidra in my experience.
you can also steer the LLM into doing tedious but correct things like injecting code, hijacking functions, intercepting data etc to gain an understanding of what's going on if straight up disassembly doesn't appear to be going in the right direction.
have the LLM periodically write reports on all its findings and then wipe the context and reset this time with the report at hand.
The “normal” Ghidra MCP is exceptionally bad but straight scripting it headless or having an LLM build a better MCP delivers similar or better results to IDA. Ghidra emulation is much more robust which is helpful. The official IDA MCP is very well made also though and it’s definitely an easier win if you want to spend the money.
I reverse engineered Civ 1 [0] to what I believe is a build sympathetic to the period. It felt a bit like software archaeology.
It took a while to get started - to begin with I think I overengineered the workflow. It was only when I rolled things back a bit and let the agents guide themselves that I started getting the best results.
They still ended up building a lot of tooling - whether or not that tooling would be useful for another decompilation project remains to be seen! MSC51 is relatively simple and the linker ordering even bounds possible variable names; so it's a pretty nice target.
[0] https://github.com/sjmelia/civd
This is why we work on making Rizin[1][2] (the engine behind Cutter[3] GUI tool) easy to use by the AI agents, both local and from the Web-interface. Because it's a tool (and a library) in pure C with very few dependencies and has well-defined (as YAML files) CLI/command interface, as well as JSON output for most of the commands, tuning it to be suitable for the AI assistants turned out to be relatively easy. Just make sure not to use the last (0.9.1) release but from current git "dev" branch, since all those improvements landed only recently and will be made into 0.10 release in 1-2 months. We also have experimental MCP server[4], but honestly, most frontier models work with CLI tools directly just fine, if not better than through MCP. Please let us know, if you meet any new (not already in issues) bug or suggestion for the improvement for such use case.
[1] https://rizin.re
[2] https://github.com/rizinorg/rizin
[3] https://github.com/rizinorg/cutter
[4] https://github.com/rizinorg/rizin-mcp
LLM assisted reverse engineering can also be a kind of "repair", meaning making things work again that have been abandoned by the original manufacturers. I did this for old audio hardware and for different CDROMs (remember them?) whose software stopped working on modern Windows, let alone Linux or Mac.
Reverse engineering with AI is really the first lever consumers have had to fight back against abusive companies since computers/software started being in everything.
It's surprising to me that anybody here is against it. Given how far things have tilted towards the companies, and that there's already legal ways to prevent piracy, it seems hard to justify vilifying it for "protect the indie games" alone. And I find it hard to believe that indie games are actually being impacted... there were a million Vampire Survivors clones before AI and I doubt anyone wants to play the slop clones anyway. The "don't reverse engineer that 25 year old game you bought so you can play it again on modern systems after we neglected to support it" are the most transparent of corporate greeds. Maybe I'm out of it.
I'm interested in AI-assisted code obfuscation. I'm particularly interested in Go but also in other languages. Has anyone experimented with that at the source code level?
How are folks getting around anthropic and openai guards against reverse engineering? They keep complaining when I try so I usually use Deepseek which doesn't seem to give a shit.
Actually, it does give a shit, in Mandarin only. In thinking it's mentioned "restrictions in Chinese language against this behavior but this is in English so it should be ok." If I remember correctly I was asking it to scan local wifi for any vulnerable devices I could put on a wall of shame at a hackathon.
In my experience with IoT and wearables, they don’t really have those. They will refuse to penetrate (root/unlock/etc), but once you’ve got access you need, you’re good to go. Except maybe if you’re adding a debug code executor (essentially a sanctioned RCE), but even that got a green light with the right wording. Heck, Fable specifically proposed clean room multi-agent processes on its own when I said “go look how stock firmware does it”, I didn’t even have to ask.
Upd: ok, I take it back, Sol just decided it doesn’t want to read a BLE capture between the host and a wearable it just uploaded our DIY firmware to. So it flashes alright, builds RCEs alright, but checking if Bluetooth HCI negotiated the good packet size is where it draws the line, huh.
I gave up and just went with DeepSeek. I have yet to hit any guardrails at all using REA.
Claude can be a prat, but ChatGPT is fine. You just have to convince them that it's for personal use. or have another model that's willing to do it start the work and then continue with a stronger model
I haven't found many issues with codex, but I'm mainly doing firmware RE, not anything to do with cybersecurity. I find even terra is so much better and faster than Deepseek for reverse engineering. The only refusal I had recently was when I asked it to add a midi sysex command to write to arbitrary ram. I got Deepseek to do that, following the pattern of the read command. Open AI models were then quite happy to use that command to upload code, debug it, etc.
I've never really hit them myself (5.5 opus, sonnet and sometimes fable).
I mostly just tell the mfs that today we're going to decompile X game because I want to improve the engine / I'm curious / I want to find out how it works. So get on with it.
This is honestly so much fun and unlocks so many different things that would have died as ideas a few years ago.
One of my favorite things that you can do is use llms to reverse engineer the harnesses they run in. Try asking claude about the tools it has access to, where it runs etc. It's surprisingly forthcoming with info, which is very helpful if you're building certain kinds of tools.
Also shout out to Matt Pocock's grill me skill, it's truly a work of art.
What is the state of software patents these days? People are barely looking at the code their agents generate, what happens when if/when infringing stuff is shipped? Do we implicitly take on liability for this when we "review" and "approve" pull requests?
This is usually among the list of reasons FOSS projects / people reject the use of LLM generated code entirely, see for example Codeberg. The answer is, it's almost certainly shipping infringing code, since there's pretty damning examples of this happening. [0]
Whether this matters is to be decided. Given that AI technologies are the backbone of the American economy right now, and given that the people running these companies are all very rich and have lots of rich friends, as well as friends in the government, and given that the world of FOSS is generally not full of very rich or very well connected people, I suspect it won't matter in the end. Some kind of carve-out will be created for LLM generated code, if necessary. Or they can just depend on the fact that nobody can afford to sue over it.
[0] https://pmc.ncbi.nlm.nih.gov/articles/PMC12377501
Who is going to de-compile/reverse ngineer GTA VI to make it run natively on "linux"?
context is always a problem as you lose too much info during the process
Itty bitty steps
where does the agent run the binary?? VM, emulator??
If you mean for embedded (the hardware devices firmware) I usually try to use or let AI make some yes VM (qemu) or emulator. Often times I even let it write custom emulator of given architecture in unicorn engine.
[flagged]
[flagged]
[flagged]
[flagged]
[flagged]
[flagged]
Just a friendly reminder that reverse engineering can have rules and regulations that are worth reviewing. For example, Apple's developer agreement specifically disallows it in certain cases.
Just do it anyway and obfuscate what you did enough that legal won’t come after you.
Companies don’t play by the rules anymore, neither should we.
Agreements can't dictate what you do in private, but they can dictate what you're allowed to submit for review. Don't submit it for review, and you'll be fine.
But if you go and edit EULA.txt to say "reverse engineering is expressly allowed", LLMs don't know the difference!