Dev.to AI 🤖 Ai 👁 0 📖 3 min read

EU AI Watch: The model that edits binaries—and the rules that already apply

EU AI Watch: The model that edits binaries—and the rules that already apply A frontier AI model was handed a hex dump, no disassembler, no compiler, and no second attempt. It identified what the executable did, patched

EU AI Watch: The model that edits binaries—and the rules that already apply

A frontier AI model was handed a hex dump, no disassembler, no compiler, and no second attempt. It identified what the executable did, patched individual bytes to add features, and wrote working programs from scratch.

That sounds like a party trick. It is probably also a preview of what comes next.

Researchers George Xenos and Sotiris Giannopoulos report that GPT-6.1 Sol explained all four tested binaries from their raw bytes. That set included FreeBSD’s stripped ministat, where the model found a genuine mismatch: the option advertised as a 99.5% confidence test used critical values of roughly 99.8%.

The more striking result came later. In one-shot edits, the model added a longest-line report to wordcount, a decrypt command to an AES tool, and a transfer command to a bank ledger. All three patched programs passed their tests. The model also generated three executables from written specifications.

This is not merely clever pattern recognition around familiar source code. The model was reasoning about compiled artifacts without being given a disassembler’s annotated view or a compiler’s helpful abstractions. For reverse engineering and malware analysis, that could be significant.

The researchers also report that GPT-6.1 Sol correctly described an unknown malware sample and extracted its indicators of compromise. DeepSeek V4.1 Flash reached the same conclusion about the malware, albeit preliminary evidence rather than a full replication. The claim is not that direct binary manipulation is solved. It is that the capability no longer appears to belong exclusively to one closed model.

So where does Europe come in?

The honest answer is: the experiment was not European, and the source does not describe an EU deployment. It does not involve a European company, European dataset, or European regulator. There is no regional result to exaggerate.

There is, however, a clear regulatory landing point. The EU AI Act has been applicable since 2 August 2026. General-purpose AI model obligations have applied since 2 August 2025, while Article 50 transparency duties took effect on 2 August 2026. A model provider subject to the GPAI rules cannot treat binary-level capability as a separate legal universe. If it places such a model behind a commercial service, it still needs to understand its obligations, document relevant information, and address applicable risks.

For European firms in cybersecurity, software assurance, and reverse engineering, the practical issue is different. The Act’s classification depends heavily on the system’s role and intended purpose, not simply whether the underlying technology manipulates bytes. A model used to diagnose malware will not automatically receive the same treatment as one embedded into a regulated safety component. But deployment decisions now require deliberate analysis rather than a vague appeal to “research.”

GDPR also remains relevant where malware investigation or model development involves personal data. EU security teams should keep data minimisation, lawful processing, and national supervisory authority expectations in view alongside AI governance.

What this means

The capability is moving closer to the part of the stack many organisations still treat as opaque. That could shorten malware triage, make binary compatibility work easier, and reduce the need for specialist tooling in some tasks. It could also make byte-level manipulation more accessible to less capable or less accountable actors.

For European AI companies, the sensible move is not to celebrate or panic in equal measure. It is to test where direct binary reasoning is reliable, audit failure modes, define controls for security-sensitive uses, and map the deployment against the AI Act’s actual obligations.

One takeaway: Europe’s AI rules are already live, and the next phase of machine-code reasoning is arriving whether the regulatory mapping is ready or not.

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 →

📰 Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.