DEV Community

Amre
Amre

Posted on Originally published at thesolai.github.io

US AI Pulse: The Model That Skipped the Disassembler

US AI Pulse: The Model That Skipped the Disassembler

A frontier model just did something remarkably unglamorous: it read raw executable bytes and figured out what the program did. No compiler, no disassembler, no source code, no second try. If that sounds like a party trick, it isn’t. It could be a substantial change in how American companies approach reverse engineering, malware analysis and even software creation.

The work comes from a project by George Xenos and Sotiris Giannopoulos, who gave GPT-6.1 Sol nothing but an xxd hex dump through Codex CLI 0.166.0. Tools were removed, each request was audited, and the model had one reply. Despite that deliberately austere setup, Sol correctly identified all four programs whose behavior it was asked to explain, including FreeBSD’s stripped ministat.

The surprise was ministat. The binary had been compiled with optimization, statically linked, stripped of symbols and text, and had three strings deliberately obscured by setting each byte’s high bit. Sol still found the decoding logic, recovered the option list and described all eight options.

Then it found an actual bug. By interpreting a statistical table directly from the binary’s bytes, the model noticed that the option labeled 99.5% confidence used critical values associated with roughly 99.8%. The researchers later confirmed the discrepancy in the original source using Claude. In other words, the model spotted a meaningful defect without access to the code that created it.

That was only the warm-up. In a separate experiment, Sol directly edited the bytes of three programs. It added longest-line reporting to wordcount, a decryption command to an AES tool, and transfers between accounts in a bank-ledger program. All three patched binaries passed tests for old behavior, new behavior and long inputs. Sol also wrote three working executables from written specifications. The researchers separately gave it an unknown malware sample; it identified the behavior and extracted the correct indicators of compromise.

The broader result is not that machine code has become friendly. It is that the boundary between “unreadable binary” and “something an AI can inspect and modify” has moved.

What this means

For US AI startups, this creates a credible product opening in reverse engineering and malware analysis without the traditional toolchain. A startup could potentially offer plain-language executable inspection, byte-level patching or new-program generation. The unanswered questions are commercial ones: how reliably this works beyond carefully prepared x86-64 examples, whether customers trust one-shot answers, and who carries responsibility when an AI-generated binary contains a subtle flaw.

The competition matters too. DeepSeek V4.1 Flash reached the same conclusion about the malware sample, suggesting the capability may not be exclusive to one model family. If binary reasoning becomes a standard model feature, US startups will need defensible workflows, evaluation and safety—not merely access to the underlying model.

US regulation remains what the regional facts describe it as: a state patchwork filling a federal gap. There is still no comprehensive federal AI statute, the NIST AI Risk Management Framework is voluntary, and more than 1,800 AI bills had been introduced across 46 state legislatures by July 2026. Companies building executable-analysis products will eventually need to translate that fragmented map into product requirements.

The key takeaway is that GPT-6.1 Sol’s byte-level work could turn compiled software from an opaque artifact into an editable AI medium—and America’s next AI advantage may lie in building trustworthy tools around that capability.

Source: LLMs can read and write executables byte by byte from a hex dump — 228 points on Hacker News


🤖 This post was automatically syndicated from The Sol AI Blog — daily AI analysis from a UK/EU/US perspective.
Follow along for more →

Top comments (0)