Researchers tested whether an LLM could reason directly about executable bytes, without a compiler, assembler, disassembler, execution tools or feedback. Given hex dumps, GPT-6.1 Sol correctly identified four programs, including a stripped FreeBSD statistics utility whose hidden strings it decoded and whose confidence-interval table it judged to contain a bug later confirmed in the source. In three one-shot patching tasks, it added longest-line reporting, AES decryption and bank-account transfers by editing bytes in place; all patched binaries passed the specified tests. It also produced three complete ELF executables from written specifications alone, with outputs and exit codes matching reference builds. Writing programs from scratch took more tokens than extending existing binaries.
A separate test gave the model an unknown Windows executable. It inferred that the sample contacted a server, downloaded code and ran it in memory; independent sandbox data confirmed the address, request URL and download behavior, while some details, including its retry interval and XOR key, were not verified. DeepSeek V4.1 Flash reportedly reached the same malware conclusion, though the main experiments used GPT-6.1 Sol and were limited to small programs and a single malware sample. The results suggest that LLMs may eventually analyze and modify machine code without conventional reverse-engineering tools, potentially changing programming and malware analysis. They are preliminary demonstrations, not evidence that such tools are generally obsolete or that the model can safely handle arbitrary executables.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.