GPT-6.1 Sol reads and writes binaries from hex dumps
Original: LLMs can read and write executables byte by byte from a hex dump
Why This Matters
Direct binary reasoning could reshape reverse engineering and malware analysis pipelines.
Researchers showed GPT-6.1 Sol can analyze, modify, and write x86-64 executables using only raw hex dumps—no disassembler, compiler, or tools—correctly explaining all five binaries tested, including real malware, in November 2026.
George Xenos and Sotiris Giannopoulos published findings showing that LLMs can now reason directly about compiled machine code—what they call "Machine Code Programming." Using GPT-6.1 Sol via Codex CLI with all tools stripped from requests, they fed the model xxd hex dumps (~2.4 tokens per byte) and asked it to do three things: explain what a binary does, extend it by editing bytes in place, and write a working executable from a description alone.
The model correctly identified all five binaries tested—including a stripped FreeBSD statistics tool called ministat (67 KB)—from raw bytes only. It also spotted a real bug in that upstream open-source tool. On a real malware sample, it correctly identified behavior and extracted valid IOCs. DeepSeek V4.1 Flash independently reached the same malware conclusions, suggesting the capability isn't exclusive to one model.
For the extension and writing tasks, outputs were scored by running them against a reference GCC build under qemu-x86_64 on Apple Silicon: stdout and exit codes had to match byte for byte. The authors frame this as a parallel to recent "Weight-Space Programming" work, where LLMs manipulated neural network weights directly. Their conclusion: future LLMs may write programs in machine code rather than high-level languages.