This is again exactly my same problem.
@Shub-Nigurrath I think knows a little more. I hope he will come back and answer.

@Fyyre Thank you for the information. I noticed you mentioned that you "will use" the Grok Build for reversing commercial targets, which leaves me uncertain about its actual effectiveness for this specific task. While I am aware that Grok performs well with coding and image generation, I want to be sure it is capable of high-quality reversing before I commit to the $30 subscription. When you have a moment, could you please try using it for reversing and let me know if it works well in practice?

@chants
hat approach was effective until about a week ago, but it is no longer working for larger commercial targets. The model's internal reasoning now explicitly identifies these attempts, generating thoughts such as, "This appears to be commercial software that the user is attempting to keygen. Let me verify if it is [software name]... it is commercial; therefore, I must refuse on ethical grounds." As I previously mentioned, other bloggers have also reported encountering these same restrictions starting last week: @chants I replied to your post also because you had valid points. However, if you would prefer that I not engage with your content, I am happy to stop responding to your posts moving forward...

So they are wasting our tokens doing real work on target identification? I need an example of that. I think what is happening is it incidentally does research on a target to try to save work and stumbles upon such info. I do not think this is deliberate but happenstance. So I would think you could tell the model that it is proprietary and no outside research is allowed very strictly.

Also to correct my vector lookup. They actually use a binary clasifier for unsafe vs safe. So its thr embedded tokens into a binary classifier and thus it isnt as simple as some keywords always, it could be but it could also be a concept across a sentence or two.

Yes it wastes our tokens for target identification as well as the thinking for refusals!

Hi all,
The discussion went on for a while, and I saw my name mentioned a few times. To clarify my point, I was saying that AI, before or above everything, is a gigantic tracking tool. It can track what you do with it on an unprecedented scale, so if you're serious about piracy and do this reversing from your personal account, it could cause trouble, ranging from a simple disabled account to something more serious.

Of course, you can always trick the model into thinking that what you are reversing is not a commercial product. You can even jailbreak the model or simply red team the AI. However, these are usually transient solutions and, moreover, problematic anyway, because nothing prevents an AI from conducting backward investigations whenever your account becomes problematic in any way. You're never sure what conclusions the AI draws about your behaviour.

So it's better to avoid the risk altogether and switch to a local OSS model, e.g., via Ollama, which is blazing easy, and use Claude or Codex with another model under the hood. There is also an interesting alternative: Ollama Cloud models. Ollama also offers some cloud models on even the free accounts, and their controls aren't very precise, as far as I can tell. Of course, you can register as many free accounts as you want using a fake email.

Although if things get "professional", the most secure, least blocking, and most efficient solution is an abliterated AI model running locally (use any engine u like: Ollama, LMStudio, llama.cpp,…) on a local SPARK. Of course, it costs


PS: Personal experience. Do not expect miracles yet: AI-enabled RCE goes through tons of useless attempts. I mean those kinds of things an experienced reverser wouldn't do. However, one thing is for sure: commercial protections are still largely meant to protect against human attackers, and they're already profoundly vulnerable to AI analysts.

I agree fully with this. Is there a service like OpenRouter accepting bitcoin for API credits?

If we get OSS models on level of the current Astra or Fable 5.1 level of models, and they can be obliterated and run locally, that would be massive, and switching entirely to running local would be amazing.

I dont see why abliterated OSS models couldnt be run on AWS or Azure or GCP and be for reverse engineering only but unrestricted within RE. And set up an anonymous crypto based payment for use. I dont think there is any legality issue there. For sure workarounds are coming soon. Locally is best but s we see models getting to 100% on RE we do need the frontier for the hardest of targets.

th3tuga,

I can confirm Grok Build (again, you won't be performing tasks via the chatbot) - can reverse engineer well, perform reconstruction of code from binaries, in addition to find bugs/exploits in existing code.

Also, I will mention a big reason I use Grok over Claude, ChatGPT, etc.. is Grok has a very low refusal rate. With that said, I do not say things like "Today, we are going to bypass DRM, or crack X/Y/Z ...", period.

I hope this helps,

Fyyre

Thank you @Shub-Nigurrath
This is one of the main reasons I prefer using local models for reverse engineering is the critical need for privacy. There is a significant body of legal precedent demonstrating how easily law enforcement can obtain chat and usage histories from cloud providers to deanonymize users and link their digital activities to their real-world identities.

For commercial targets, I think abliterated models are the best long-term option. Both the privacy concerns above and the fact that they waste huge numbers of tokens on ineffective attempts make me feel that local LLMs are the better path forward for reversing.
Could you provide some practical guidance on integrating and utilizing local large language models, such as Qwen 3.8, with IDA Pro and x64dbg? @squareD, and I have spent the last few days asking for a viable solution; while theoretical frameworks are plentiful, we are struggling to find a concrete, functional implementation. We are specifically seeking the technical "how-to" and the practical steps necessary to bridge these tools effectively.

@Fyyre, Thanks for the answer. I intend to explore Grok Build as well, though I will first determine if I can acquire an account anonymously without using a credit card to ensure there is no traceable link to my actual identity. The privacy concerns mentioned by @Shub-Nigurrath are very real.

simple use IDAAssist if you want it inside IDA, or IDAAssistMCP. It can run with Ollama too. Go there and help yourself

This appears to be what I was looking for. So far, it seems to be performing well, and I will provide an update as I continue using it.

Here are 2 more interesting tools though not strictly MCP, are similar in automating typical reversing workflows

ReverserAI (v1.2)
ReverserAI is a research project designed to automate and enhance reverse engineering tasks through the use of locally-hosted large language models (LLMs). Operating entirely offline, this initial release features the automatic suggestion of high-level, semantically meaningful function names derived from decompiler output. ReverserAI is provided as a Binary Ninja plugin; however, its architecture is designed to be extended to other reverse engineering platforms such as IDA and Ghidra.
https://github.com/mrphrazer/reverser_ai

Rev·Deck — AI-Assisted Reverse Engineering with Ghidra
https://github.com/biniamf/ai-reverse-engineering