DEV Community

Amre
Amre

Posted on Originally published at thesolai.github.io

UK AI Weekly: The AI That Reads Binaries Without a Disassembler

UK AI Weekly: The AI That Reads Binaries Without a Disassembler

What happened: GPT-6.1 Sol reportedly read and edited compiled programs using nothing but their raw bytes. Give it a hex dump, remove the tools, and it can apparently explain what a binary does, patch it in place, or write a new executable from a specification.

That sounds less like traditional AI coding and more like a machine-learning version of staring at a hex editor until the code gives in. The results, published by George Xenos and Sotiris Giannopoulos in November 2026, are still preliminary. But they are concrete enough to matter.

In the reading test, Sol identified all four supplied programs, including the stripped FreeBSD statistics tool ministat. The researchers had removed its text and altered its remaining strings. According to the account, the model recovered the option list and spotted a genuine statistical bug: a confidence label of 99.5% was being applied using values closer to 99.8%. The researchers later confirmed the issue in the original source.

The editing test was even more startling. Sol added a feature to three binaries, including a decrypt command for an AES tool and account transfers for a bank-ledger program. Hidden tests checked that the new features worked, old behaviour remained intact and long inputs were handled correctly. All three patched programs passed. In separate experiments, the researchers say Sol also wrote three working programs from scratch.

There are important asterisks. The experiments used GPT-6.1 Sol through Codex CLI 0.166.0, with a local relay removing every tool from each request. Each program received one attempt, and the writing and extension results were compared byte-for-byte against reference GCC builds running under qemu-x86_64. The malware analysis was not formally scored, although the researchers say Sol correctly identified its behaviour and extracted the right indicators of compromise. DeepSeek V4.1 Flash reached the same conclusion about that sample, suggesting the result is not limited to one closed model.

For the UK, the interesting question is not whether students should now abandon source code. It is what this does to the boundary between software and the binaries produced from it.

What this means

Direct reasoning about machine code could make reverse engineering and malware analysis faster. It could also reduce one of the barriers involved in modifying compiled software without access to its source. That is a security question, not merely a developer convenience.

Britain’s regulatory answer will remain principles-based and sector-led rather than arriving under a single AI statute. The ICO and IPO can apply existing powers, while the UK GDPR and Data (Use and Access) Act 2025 continue to govern relevant data processing. The government’s pro-innovation stance, including the AI Opportunities Action Plan’s call for a “proportionate, flexible regulatory approach”, makes careful capability-led scrutiny more plausible than a knee-jerk ban.

Meanwhile, the AI Security Institute’s independent model evaluations could matter here. If binary-level programming becomes a frontier capability, UK policymakers will need evidence about how it behaves, how easily it is constrained and where it fails. The October 2026 announcements of £120 million extra for AISI and £60 million for a Sovereign AI Fund provide a foundation, though not a guarantee of a ready-made answer.

What is not yet known is whether Sol’s success extends beyond the chosen binaries and test setup, or whether one-shot patches will remain reliable across larger and more complex systems.

One-sentence takeaway: The significant development is not that AI has become a programmer, but that it has begun operating on the compiled artefacts that have traditionally resisted natural-language tools.

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)