Being able to hook Claude code up to this has made reversing way more productive. Highly recommend!
It works surprisingly nicely with AI agents (I mean, like Cursor or Claude Code, I don't let it run autonomously!).
Here on detecting malware in binaries (https://quesma.com/blog/introducing-binaryaudit/). I am now in process of recompiling and old game Chromatron, from PowerPC binary to Apple Silicon and WASM (https://p.migdal.pl/chromatron-recompiled/, ready to play, might be still rough edges).
As well as the research history (slated to be updated in a few days): https://mahaloz.re/dec-progress-2024
I've used IDA, Ghidra, and Binary Ninja a lot over the years. At this point I much prefer Binary Ninja for the task of building up an understanding of large binaries with many thousands of types and functions. It also doesn't hurt that its UI/UX feel like something out of this century, and it's very easy to automate using Python scripts.
Binary Ninja – an interactive decompiler, disassembler, debugger - https://news.ycombinator.com/item?id=41297124 - Aug 2024 (1 comment)
Binary Ninja – 4.0: Dorsai - https://news.ycombinator.com/item?id=39546731 - Feb 2024 (1 comment)
Binary Ninja 3.0: The Next Chapter - https://news.ycombinator.com/item?id=30109122 - Jan 2022 (1 comment)
Binary Ninja – A new kind of reversing platform - https://news.ycombinator.com/item?id=12240209 - Aug 2016 (56 comments)
Can't speak to this as I don't RE for security purposes, but:
> no plugin support and rather limited IR.
this I'm profoundly confused by. BN has multiple IRs that are easily accessible both in the UI and to scripts. And it certainly has a plugin system too.
We have actually been more inspired by Jetbrains lately than VS Code. Take that for what you will.
We do try to pick simple sane defaults while still allowing enough customization to adapt to different workflows.
Actually working on a startup wizard for first time users if they want to more closely replicate the feel of other RE tools since muscle memory is hard to break.
https://github.com/Vector35/binaryninja-api/releases/downloa...
I guess one issue I have is that I don't have good ideas of fun projects, and that's probably something I need to actually get the motivation to learn. I can find a "hello world", that's easy, but it won't help me get an idea of what I could reverse engineer in my life.
For instance I have a smartspeaker that I would like to hack (being able to run my own software on it, for fun), but I don't know if it is a good candidate for reverse engineering... I guess I would first need to find a security flaw in order to access the OS? Or flash my own OS (hoping that it's a Linux running there), but then I would probably want to extract binary blobs that work with the buttons and the actual speaker?
The good news is that there has never been MORE resources out there. If you want to use this learning expedition as an excuse to also build up a small electronics lab then $100 on ali express to buy whatever looks cheap and interesting and then tear it apart and start poking around to find where the firmware lives. Pull the firmware, examine it, modify it and put it back :)
This guy has a discord server with a specific "book club" section where they all choose a cheap $thing and reverse engineer it: https://www.youtube.com/@mattbrwn/about
I can't help much with "traditional" app/software RE work, sorry.
Thanks a lot!
Turns out that frontier grade llms are absolutely fantastic for extremely advanced static analysis. If you go one step further and manage to get your firmware running inside of an emulator or other place where you can attach GDB... Then putting an mCP server on that as well unlocks so much insane potential.
I feel like the tendency for people to assume others have nearly $500 or so of credits on their AI to blow every month is kinda crazy.
Reminds me of the "just get Netflix, Prime, etc." ending up with a $100/m bill.
I can't comment on the self-hosted models but if they support tool calling / MCP then it should work. How well may be a function of model size/arch/quant ... etc though.
> I feel like the tendency for people to assume others have nearly $500 or so of credits on their AI to blow every month is kinda crazy.
My use is hobby level so $20/mo chatGPT sub is enough. I do hit my 5hr limits on a regular-ish interval but that's usually a sign that I need to step away from the work anyways.
The nightmare course explicitly talks about how to use Ghidra.
1: https://guyinatuxedo.github.io 2: https://www.roppers.org
from then you can use things like Ghidra (which supports a lot of those old CPU arches) for more advanced analysis and make the game do almost whatever the hell you want if you have the patience.
I think a lot of the skills will transfer quite well (obviously not 1:1, you will need to learn some things) to the more employable side of RE if that's what you're interested in
I guess I'm struggling to transfer that to "real-life" scenarios. Like getting something useful out of reverse engineering (getting infinite lives is interesting to see that I can tamper with the game, but it's not exactly useful).
(Thinking more of license-checking, and serial-number generation rather than infinite lives.)
One thing which really helped me (and I wholeheartedly recommend) is to write simple programs, run them through the compiler and then in the disassembler. It really helps build a correspondence between program structure and its object code.
Eventually, you can make it even more fun and challenging by stripping debug symbols and turning on compiler optimisations.
Happy reversing!
The book is designed for beginner and advance users.
The reverse engineering I've learned has generally been to fix something that has annoyed me - for example I reverse engineered part of RCT3 to fix mouse input with high poll rates and allow for resizable windows [0]. Certainly easier to approach than trying to get into a closed device since you can attach a debugger.
Then figure out what SoC your smart speaker is running.
A lot of soc vendors just provide SDKs for IoT applications. Maybe you'll find something like that you can flash.
Also if there's a separate SPI flash chip, you can simply dump it and study the firmware.
So for the second thing, pulling the data off chips like that typically involves some specialized hardware, and you have to potentially deal with a bunch of cryptographic safeguards to read from the chip’s memory. Not impossible though, and there are not always good safeguards, but might be worth checking out some simpler programs and working up to it, or learning some basic hardware hacking to get an idea of how that process works.
Oh I’m sorry the NSA didn’t spy on the whole country “wittingly” according to our leaders, carry on and use their software no ethical conflict here.
Who built the Echelon follow-up, proto-dragnet system that provided the framework for the spying you bemoan? — the one extended and taken live in the early 2000s? Those same 90s hackers you glorify.
The community I’m talking about definitely weren’t like secretly building tools for these agencies. I mean this sincerely I have no idea what point you’re making. The agencies existed and made tech so by logical necessity people worked there. I didn’t say all people in computing.
There was a prevalent community of programmers and hackers who understood what these organizations represented and would never be on a forum blithely talking about some tool they made as if it was acceptable. Shame on anyone using these tools and the lack of objection to this post is a metric of how disgusting computing culture and really this forum are.
You knew exactly what point I’m making, because it’s the first thing you responded to. And indeed, what you responsed to throughout your question. So no, you’re not being sincere.
Those groups always interacted and your bald assertion of their morality is directly contradicted by my experience of their interactions (eg, criminals and government corresponding at UW) and the change in Boomer and Gen X hackers following 9/11.
> There was a prevalent community of programmers and hackers who understood what these organizations represented and would never be on a forum blithely talking about some tool they made as if it was acceptable.
From their computers that originated in a US Navy lab?
Again, my experience from Seattle is that the idealism was always more show than reality — and government technologies were not only consumed, but built on contract when interests aligned (eg, stopping cyber warfare or dismantling terrorist networks).
What you’re describing — ineffective moral absolutism — wasn’t what I recall from the 90s hacker ethos that always existed in a liminal zone, but rather the 2010s era co-opting of existing groups (eg, Anyonymous) for moral crusading.
This is a logical non sequitur. It’s like being fed your lunch by your kidnapper and when you protest they say, you’re using the energy I provided to protest. Like that’s a contradiction somehow, it’s not.
Whatever the participation in the past may have been does nothing to excuse or make okay future behavior.
And to the extent I did understand your point, I was confused because it was such a strawman. As I explained I wasn’t talking about every person in computing (of course). Interacted doesn’t mean they were the same or had no moral distinction.
> ineffective moral absolutism
I don’t even know where to start with how flawed your thinking is. Effectiveness isn’t the driver for having morals. And obviously it was effective because I’m here protesting. Neither is it absolutism. Objecting to the gross abuse of our government doesn’t equate with absolutism.
Computers should serve their owners not corporate interests and dragnet surveillance. That was understood. 2010s Anonymous was a different thing in a different context that I wasn’t contemplating.
what sort of device did you type this comment from?
However you slice it you have no argument here. Being held hostage by corporate interests in computing doesn’t make it okay to hold hands with one of the worst organizations on Earth.
By your twisted logic, because it’s practically impossible to avoid proprietary firmware owned by corporations, we should completely acquiesce to all forms of digital abuse by the government and actually should participate in their computing programs. Listen to yourself.
You’d be telling people who lived under an evil dictatorship to shut up because they walk the roads the govt provides.
I once tried learning how to RE with radare2 but got very frustrated by frequent project file corruption (meaning radare2 could no longer open it). The way these project files work(ed?) in radare2 at the time was that it just saved all the commands you executed, instead of the state. This was brittle, in my experience.
I don't have a lot of free time, so I have to leave projects for long periods of time, not being able to restart from a previous checkpoints meant I never actually got further.
IIUC, one of the first things Rizin did was focus on saving the actual state, and backwards/forwards-compatibility. This fact alone made me switch to Rizin. To its credit, my 3-year old project file still works!
Now for the downside: there is apparently a gap in Windows (32-bit) PE support, causing stack variables to be poorly discovered: https://github.com/rizinorg/rizin/issues/4608. I tested this on radare2, which does not have this bug. I'm hoping this gets fixed in Rizin at some point, at which point I'll continue my RE adventure. Or maybe I should give an AI reverse engineer a try... (https://news.ycombinator.com/item?id=46846101).
(Funny expression, that. I'll wait, of course. It'll be a happy day when this works again and I can slowly make progress RE'ing again.)
Your corruption frequency anecdote matches mine. I don't have the mental werewithal to deal with that. I won't go back to radare2 until they change their project file stability somehow.
It's a relocatable object file exporter that supports x86/MIPS and ELF/COFF. In other words, it can delink any program selection and you can reuse the bits for various use-cases, including making new programs Mad Max-style.
It carved itself a niche in the Windows decompilation community, used alongside objdiff or decomp.me.
> The relocation table synthesizer analyzer relies on a fully populated Ghidra database (with correctly declared symbols, data types and references) in order to work
It's a shame that this requirement exists (I am well aware that it's a functional necessity), because all the stuff I want to relink is far too big to make a full db!
Even then, a full DB is quite achievable, even on large projects. The biggest public project using ghidra-delinker-extension out there is the FUEL decompilation: https://github.com/widberg/FUELDecompilation
The executable is 7 MiB, has over 30,000 functions and has more than 250,000 relocations spots. The user made the game relocatable in six weeks (with four of them debugging issues with my extension). They then managed to replace code in spite of the fact that the artifact was built with LTO by binary patching __usercall into MSVC.
There's a write-up about all of that that is well worth a read: https://github.com/widberg/fmtk/wiki/Decompilation
I've also had one user manage to fully delink the original Halo on the Xbox in one week. To be fair, they were completely nerd-sniped and worked non-stop on it, but it still counts.
I might try a partial delink and see how it goes!
The Windows decompilation community is far more fragmented than the console one, as it hasn't coalesced around a common set of tools like splat or decomp-toolkit.
I've done a case study where I've ported a Linux a.out program into a native Windows PE program without source code: https://boricj.net/atari-jaguar-sdk/2023/11/27/introduction....
Another case study was ripping the archive code from a PlayStation game and stuffing it into a Linux MIPS program to create an asset extractor: https://boricj.net/tenchu1/2024/03/18/part-6.html
For embedded IDA is very ergonomic still, but since it’s not abstract in the way Ghidra is, the decompiler only works on select platforms.
Ghidra’s architecture lends itself to really powerful automation tricks since you can basically step through the program from your plugin without having an actual debug target, no matter the architecture. With the rise of LLMs, this is a big edge for Ghidra as it’s more flexible and easier to hook into to build tools.
The overall Ghidra plugin programming story has been catching up; it’s always been more modular than IDA but in the past it was too Java oriented to be fun for most people, but the Python bindings are a lot better now. IDA scripting has been quite good for a long time so there’s a good corpus of plugins out there too.
Ghidra is the better tool if you're dealing with exotic architectures, even ones that you need to implement support for yourself. That's because any architecture that you have a full SLEIGH definition for will get decompilation output for free. It might not be the best decompiler out there, sure, but for some architectures it's the only decompiler available.
Both are generally shit UX wise and take time to learn. I've mostly switched from IDA to Ghidra a while back which felt like pulling teeth. Now when I sometimes go back to IDA it feels like pulling teeth.
- AVR
- Z80
- HC08
- 8051
- Tricore
- Xtensa
- WebAssembly
- Apple/Samsung S5L87xx NAND controller command sequencer VLIW (custom SLEIGH)
It's also not about lack of support, but the fact that you have to pay extra for every single decompiler. This sucks if you're analyzing a wide variety of targets because of the kind of work you do.
IDA also struggles with disasm for Harvard architectures which tend to make up a bulk of what I analyze - it's all faked around synthetic relocations. Ghidra has native support for multiple address spaces.
Maybe we need to get some good cracked^Wcommunity releases of Binja so that we can all test it as thoroughly as IDA. The limited free version doesn't cut it unfortunately - if I can't test it on what I actually want to use it for, it's not a good test.
(also it doesn't have collaborative analysis in anything but the 'call us' enterprise plan)
(not if you're only doing x86/ARM stuff, though)
I was recently trying to analyse a 600mb exe (denuvo/similar). I wasted a week after ghidra crashed 30h+ in multiple times. A seperate project with a 300mb exe took about 5h, so there's some horrible scaling going on. So I tried out Ida for the first time, and it finished in less than an hour. Faced with having decomp vs not, I started learning how to use it.
So first difference, given the above, Ida is far far better at interrupting tasks/crash recovery. Every time ghidra crashed I was left with nothing, when Ida crashes you get a prompt to recover from autosave. Even if you don't crash, in general it feels like Ida will let you interrupt a task and still get partial results which you might even be able to pick back up from later, while ghidra just leaves you with nothing.
In terms of pure decomp quality, I don't really think either wins, decomp is always awkward, it's awkward in different ways for each. I prefer ghidra's, but that might just be because I've used it much longer. Ida does do better at suggesting function/variable names - if a variable is passed to a bunch of functions taking a GameManager*, it might automatically call it game_manager.
When defining types, I far prefer ida's approach of just letting me write C/C++. Ghidra's struct editor is awkward, and I've never worked out a good way of dealing with inheritance. For defining functions/args on the other hand, while Ida gives you a raw text box it just doesn't let you change some things? There I prefer the way ghidra does it, I especially like it showing what registers each arg is assigned to.
Another big difference I've noticed between the two is ghidra seems to operate on more of a push model, while Ida is more of a pull model - i.e. when you make a change, ghidra tends to hang for a second propagating it to everything referencing it, while Ida tries pulling the latest version when you look at the reference? I have no idea if this is how they actually work internally, it's just what it feels like. Ida's pull model is a lot more responsive on a large exe, however multiple times I've had some decomp not update after editing one of the functions it called.
Overall, I find Ida's probably slightly better. I'm not about to pay for Ida pro though, and I'm really uneasy about how it uploads all my executables to do decomp. While at the same time, ghidra is proper FOSS, and gives comparable results (for small executables). So I'll probably stick with ghidra where I can.
During the startup auto analysis? For large binaries it makes sense to dial back the number of analysis passes and only trigger them if you really need them, manually, one by one. You also get to save in between different passes.
I figured I probably could remove some passes, but being a lite user I don't really know/didn't want to spend the time learning how important each one is and how long they take. Ida's defaults were just better.
If it was mine, I'd stick in a heavily-obfuscated bit of code for people to use Ghidra to reverse-engineer Ghidra to find out what it did.
OMG, it's a keylogger! But it's not activated by default! How does it get activated? How does it emit its logs?
And after weeks of poking they find you run it with "$ ACTIVATE_KEYLOGGER=true ghidra", and it saves to "$HOME/Desktop/ghidra_key_log.txt"
Haha, psych!
https://www.youtube.com/watch?v=d7qVlf81fKA&list=PL4X0K6ZbXh...
(#3 forward uses Ghidra)
It worked fine in Ubuntu and Windows. The interface takes some getting used to, but paired with Bless Unofficial (using snap to install), it makes reverse engineering smooth.
I was a special agent with an org involved in similar work. They put me through 7 SANS courses, including paying for 5 certs, in 18 months.
It's difficult to be an AI doomer when you see stuff like this.
Ghidra excels because it is extremely abstract, so new processors can be added at will and automatically have a decompiler, control flow tracing, mostly working assembler, and emulation.
IDA excels because it has been developed for a gazillion years against patterns found in common binaries and has an extremely fast, ergonomic UI and an awesome debugger.
For UI driven reversing against anything that runs on an OS I generally prefer IDA, for anything below that I’m 50/50 on Ghidra, and for anything where IDA doesn’t have a decompiler, Ghidra wins by default.
For plugin development or automated reversing (even pre LLMs, stuff like pattern matching scripts or little evaluators) Ghidra offers a ton of power since you can basically execute the underlying program using PCode, but the APIs are clunky and until recently you really needed to be using Java.
I think what NSA is likely to keep confidential are in-house plugins that are so specialized and/or underengineered that their publication would give away confidential information: stolen and illegitimate secrets (e.g. cryptographic private keys from a game console SDK), or exploits that they intend to deny knowledge of and continue milking, or general strategies and methods (e.g. a tool to "customize" UEFI images, with the implication that they have means to install them on a victim's computer).
They create executables, which contain encrypted binary data. Then, when the executable runs, it decodes the encrypted data and pipes it into "sh".
The security is delusional here - the password is hard coded in the executable. It was something like "VIVOTEK Inc.".
Ghidra was able to create the C code and I was able to extract also the binary data to a file (which is essentially the bash script).
The password would be visible, but the encyption algorithm and the script’s text wouldn’t.
(Btw, these links are just for anyone curious to read more - reposts are fine after a year or so - https://news.ycombinator.com/newsfaq.html)
NSA Ghidra open-source reverse engineering framework - https://news.ycombinator.com/item?id=40508777 - May 2024 (61 comments)
Ghidra 11.0 Released - https://news.ycombinator.com/item?id=38740793 - Dec 2023 (11 comments)
Ghidra 10.3 has been released - https://news.ycombinator.com/item?id=35908418 - May 2023 (6 comments)
NSA Ghidra software reverse engineering framework - https://news.ycombinator.com/item?id=35324380 - March 2023 (103 comments)
Ghidra: Software reverse engineering suite developed by NSA - https://news.ycombinator.com/item?id=33226050 - Oct 2022 (42 comments)
Ghidra: A software reverse engineering suite of tools developed by the NSA - https://news.ycombinator.com/item?id=27818492 - July 2021 (142 comments)
Ghidra 9.2 - https://news.ycombinator.com/item?id=25086519 - Nov 2020 (78 comments)
The Ghidra Book - https://news.ycombinator.com/item?id=24879314 - Oct 2020 (5 comments)
Ghidra Decompiler Analysis Engine - https://news.ycombinator.com/item?id=19599314 - April 2019 (30 comments)
Ghidra source code officially released - https://news.ycombinator.com/item?id=19572994 - April 2019 (7 comments)
Ghidra Capabilities – Get Your Free NSA Reverse Engineering Tool [pdf] - https://news.ycombinator.com/item?id=19319385 - March 2019 (17 comments)
Ghidra, NSA's reverse-engineering tool - https://news.ycombinator.com/item?id=19315273 - March 2019 (405 comments)
Ghidra - https://news.ycombinator.com/item?id=19239727 - Feb 2019 (59 comments)
NSA to Release Their Reverse Engineering Framework GHIDRA to Public at RSA - https://news.ycombinator.com/item?id=18828083 - Jan 2019 (90 comments)
It's certainly not the first thing they've released (selinux, for one, and then all the other repos in the account), but this repo showing up on HN, with a prominent call-to-action to look at a career with them, is a great way to target the applicants you want ("those who would find this project interesting, because it's just the sort of thing we need them to work on")
Atlassian used to do (maybe still does) this in bitbucket if you open dev tools - a link to their careers page shows up
when i try to expand their faq, it seem to try an open a (presumabl) malicious link , i wont paste the link here just in case it is really malicious
I wonder what is the purpose of ghidralite dot com. SEO spam? Are they building trust and then will swap out the Download button with a poisoned binary.
amazing tool
This is indicative of two things.
1. While I can't stand the guy, ya'll need to watch Peter Thiel's talk from 10-15 years ago at Stanford about not building the same thing everyone else is, a la, the obvious thing.
2. People are really attracted to using LLMs on deep thinking tasks, off shoring their thinking, to a "Think for me SaaS". This won't end well for you, there's no shortcuts in life that don't come with a (huge) cost.
The person who showed their work and scored A's on math tests instead of just learning how to use a calculator, is better off in their career/endevours than the 80% of others who did the latter. If Laurie Wired makes an MCP for Ghirda and uses it that's one thing, you using it without ever reverse engineering extensively is completely different. I'd bet my bottom dollar that Laurie Wired doesn't prefer the MCP over her own mental processes 8/10 times.
It's useful for the automation of small repetitive tasks here and there. I was never expecting it to gain the traction that it did; anyone saying they expect it to replace reverse engineers (it won't) is wildly misunderstanding the original intent.
Quite trivial to create binaries that massively confuse LLMs!
I, too, watched The Sorcerer’s Apprentice. The problem is that I, too, shipped a fuckton of working, reviewed, reworked, tests-and-lint-passing, properly-typed code implementing brand new features from scratch in the last 48 hours, that would have taken me 48 days a year ago.
“thinking” means a lot of different things, and you can indeed outsource a lot of it to other things that can think at different levels of ability than you. This is effectively what an engineering organization does.
Perhaps I haven’t fully offshored my thinking in the sense you mean in that I review all the code and give feedback on the PRs—I still steer. But I think the SOTA will continue to improve until we can indeed oneshot larger and larger tasks.
I was talking about critical tasks where human nuance is important, just because an LLM can produce a result, does not mean that the result is great. Not everything people work on are "features" delivered via http handlers.
I don't understand this new paradigm where everyone wants to brag about how quick they get X amount of work done. Its the long standing belief of pretty much any quality builder that quick != quality, and quick usually isn't necessary. I'm glad your KPIs are great though and your product is getting 2 months of features every two days... The world needs this!
The short summary of it is: the sorcerer's apprentice (Mickey) uses magic to get a broom to fetch water for him, and then the situation gets out of control as the broom continues to get water, and he has no idea how to stop it.
(It's a cautionary tale about the danger of playing with forces you don't really understand/"be careful what you wish for".)
No KPIs, just useful tools for myself and anyone else who finds them valuable.
Obviously I get much less out of that — I’m not denying the tradeoff, just saying that some people are all the way to “write a short request, accept the result” for (certain) thinking tasks.
That is, we seem to be conflating different cases - ie, being an expert versus hiring an expert. A manager and an SDE get different utility from the LLM.
I think I expressed it poorly, but I think that we need to consider that outsourcing thinking entirely is the right answer in the way that subcontracting or outsourcing or hiring itself can be; and that we seem to get caught in a “spectrum” or false dichotomy (ie, “is outsourcing good or bad?”) discussion, when the actual utilization of LLMs, their content, etc interacts in a complex way due to the diversity of roles, needs, etc that humans themselves have. And the impact on acquired expertise is only one aspect, for which “less work, less learning” is both true but too simple.
Are you just trying to argue for the sake of arguing?
Being first and the winner requires a lot to line up, so it shouldn't be the only, default, or best setting. Pursuing this is optimizing.
Also a message from 10-15 years ago might not reflect the same context as today.
In other words, what's good for Peter Theil might not be goid for you.
Edit: Wikipedia has a table with 1.0 being 2003 https://en.wikipedia.org/wiki/Ghidra