thro 21 Yeah but the ai doesn’t need the ascii assembly representation. Theyre literally just for humans and maps 1:1 to the bytecode.
The ai doesn’t need to disassemble because it understands the binary opcodes.
It’s not that simple because x86 has alignment so just pasting hexadecimal doesn’t confirm that this is even an application. I mean there’s a process because you’ve got headers, you have an import table, you have an export table. I mean software actually has a framework for it to exist. That’s why I said you don’t understand what you’re saying. I am giving you a token representation because I do understand what you’re asking about doing and I’m explaining to you how the tokens would be laid out in order to accomplish something like what you are asking.
Without a framework of understanding, raw bytecode could be an image, it could be a document, it could be a data file, it could be an archive. There’s no way to specifically know what data it is unless it’s marked and explained.
Currently, there is no AI model that 100% understands raw hexadecimal bytecode as entire applications without any tools. There are tools that run which the model calls to run different softwares in order for the models to inspect different things, and that is already complicated on the reverse engineering side because of the sheer number of tokens that can be absorbed from basically nothing with it just reading through different bytecodes. So managing context window sizes is already a challenge if you want to use LLMs for reverse engineering today, and I know that there’s been a lot of development in this front in the last year in terms of guiding the LLM as to what it should inspect and guiding what information should be in the context window and not running a single context window, but at the same time, we’re still not anywhere near what you’re asking.
1 Like
thro 23 Meanwhile it’s already working, with tool use.
Calvin 24 I’d really love some details on how he set that up.
i.e. prompts, harness details, tool setup, etc.
That stuff can be overwhelming and I have yet to find a good tutorial on when and how to use this stuff. Turnstone seems amazing but just using it bare was giving me worse results than built in coding clients like Zoo Code.