GPT-6.1 Sol reads and writes x86-64 executables byte by byte from a hex dump
- The authors report that GPT-6.1 Sol read and wrote x86-64 executables byte by byte from a hex dump, with no disassembler, compiler, or other tools.
- Given only the bytes, it explained all five binaries it was given, including a stripped open-source statistics tool, and found a real bug in that tool's upstream code.
- It correctly identified the behavior and extracted the correct IOCs from an unknown malware sample, and DeepSeek V4.1 Flash reached the same conclusion about that malware.
- In a later experiment it added a feature to three binaries by directly editing their bytes and wrote three programs from scratch.
Hacker News opinions
Wow, it can read raw bytes like natural language? The impressive part is that it edits bytes directly without going through ASM, which normally means a human lifting bytes to ASM and back, and that takes a lot of knowledge.
I've suspected for years that AI companies trained on commercial software like games, audio plugins, and DAWs. Wouldn't surprise me if they used torrented pirated apps like they did with books and music. How else do you explain the avalanche of vibe coded audio plugins this year?
Vibe coding only became viable last year, since Opus 4.5 came out Nov 24, 2025. Plugins are a thing lots of people wanted to try, and now they can, so you get the flood.
Binaries are basically the pure logic of the code without the identifiers, and vendors ship tons of them. Battle-tested code you'd never get your hands on otherwise. Training on it makes perfect sense to me.
Not sure about that. AIs are quite good at using existing binary analysis tools, and they usually don't operate on binaries in prompt space like this.
You don't need an LLM to disassemble 6502 code. Disassembly has been a solved problem for decades.
This is useful for reverse engineering and malware analysis where there's machine code and no source. But generating machine code directly instead of high level languages doesn't make sense to me. LLM output is non deterministic, so it's better to generate source once and compile it with a deterministic compiler for each target.
We also have more or less deterministic tools to lift code, and those are usually more efficient to use.