- LLMs can now understand and write machine code.
- Given only a hex dump, with no disassembler or compiler, and no tools, GPT-6.1 Sol can read and write executables byte by byte. This is an early sign that LLMs can program directly in binary.
- By reading only the actual bytes, it explained all five binaries we gave it, including a stripped open-source statistics tool, where it even found a real bug in the upstream code.
- It successfully analyzed an unknown malware sample we gave it, correctly identifying its behavior and extracting the correct IOCs.
- In a later experiment, it added a feature to three binaries by directly editing their bytes. It also successfully wrote three programs from scratch.
- In the future, LLMs may write directly in machine code instead of using high level languages.
- As preliminary evidence that this isn't limited to closed models, DeepSeek V4.1 Flash reached the same conclusion about the malware.
- These findings raise interesting questions about the future of reverse engineering, malware analysis and even programming.
Read full story: https://sgiannopoulos.github.io/blog/machine-code-programming/