---
title: 'In the News: November 21, 2026 (Evening)'
description: 'With no compiler or assembler, GPT-6.1 Sol patched and wrote small executables byte by byte, and each matched a reference build on every test.'
canonical_url: 'https://darkfactory.dev/news/2026-11-21-evening'
markdown_url: 'https://darkfactory.dev/news/2026-11-21-evening.md'
collection: news
date_published: '2026-11-21T19:01:00-05:00'
date_modified: '2026-11-21T19:01:00-05:00'
---

# In the News: November 21, 2026 (Evening)


Two researchers stripped every tool from a frontier model, handed it hex dumps, and asked it to change and write compiled programs directly in bytes. Every patched and written program matched a reference build on every test case. The sample is small and each task got one attempt, so the capability claim is early. The way they scored it is usable now.

## 1. A model with no toolchain patched and wrote working executables from a hex dump

**[Programming in the Machine Code Space: LLMs can read & write executables byte by byte](https://sgiannopoulos.github.io/blog/machine-code-programming/)** · George Xenos and Sotiris Giannopoulos · Blog post, November 2026

The authors ran GPT-6.1 Sol at high effort through Codex CLI, with a local relay removing every tool from every request: no compiler, assembler, disassembler or execution. The setup notes put the rule plainly: "One reply: no feedback and no second try." Given only hex, the model described all four programs it was shown, including FreeBSD's ministat built at -O3, statically linked, stripped and with its last three strings hidden. Reading ministat's t-table as raw floats, it reported that the option labeled 99.5% confidence uses critical values for about 99.8%. The authors say they then confirmed that bug in the upstream source.

The scored experiments carry more weight. Asked for one new feature per program, the model replied with same-length byte edits: 74 bytes (1.4% of the file) to make a word counter report its longest line, 544 bytes to add AES decryption, and 379 bytes to add a transfer command to a bank ledger. Asked to write the same three programs from a written spec, it produced complete executables of 391, 4,144 and 3,584 bytes, headers included. For the patches, hidden tests checked old behavior, the new feature and long inputs. Each build ran under qemu-x86_64 on every test case alongside a reference GCC build, and stdout and exit code had to match byte for byte. The authors report that all six passed every test, with one attempt per program.

The limits are stated in the post. A hex dump costs about 2.4 tokens per byte, so the 67 KB ministat took about 163,000 tokens, and the authors say they do not know how quality holds up on larger programs. The descriptive runs and a malware analysis were not formally scored. In that malware run, on a 4,608-byte Windows sample, the model's report matched a public sandbox analysis on the server address and request URL. A repeat with the open-weight DeepSeek V4.1 Flash reached the same conclusion using about 17 times as many reasoning tokens, and named 9 of 13 hidden API functions correctly. For that repeat, the authors say they could not verify which tools, if any, the hosted provider's request included. The post had 178 points and 119 comments on [Hacker News](https://news.ycombinator.com/item?id=50546587) at 19:02 EST on November 21, after it was submitted on November 18.

**Why it matters:** The patched and written programs were judged the way output nobody can read has to be judged: run side by side with a reference build on a fixed test set, with exact matching on output and exit code. That is the acceptance check to copy when an agent produces something no reviewer will open. Treat the capability itself as an early result, since it rests on one attempt each on programs under 10 KB.
