Research shows LLMs can parse and modify compiled executables from raw hex dumps
A study tests whether general-purpose LLMs can reason directly about compiled binaries' bytes, without compilers, disassemblers or execution—an approach the researchers call Machine Code Programming. Given hex dumps alone, the model correctly identified four test programs, including a stripped, statically linked FreeBSD binary (ministat) whose strings had been obfuscated. In the process, the model flagged a real statistical bug in ministat's confidence-interval table, which the researchers say they verified against the original source code.
GoKawiil's interpretation of the reporting above, not reported fact.
If LLMs can reliably interpret and manipulate raw machine code, this could open new possibilities for security analysis, malware reverse engineering, and binary patching without relying on traditional disassembly toolchains. The researchers frame this as parallel to prior work showing LLMs can directly manipulate neural network weights, suggesting a broader pattern of models handling 'unreadable' binary artifacts. However, this appears to be an early demonstration on a small set of examples, so its reliability at scale remains unproven.
- LLMs successfully identified four programs purely from hex-dump bytes, with no compiler or disassembler involved.
- The model decoded obfuscated strings and reconstructed a full options list from a stripped, statically linked FreeBSD binary.
- The LLM discovered a genuine bug in ministat's confidence-level table directly from bytes, later confirmed in the source code.
Source: sgiannopoulos.github.io, 2026-11-18
Published there as: “LLMs can read and write executables byte by byte from a hex dump”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.