OpenAI GPT-6 Astra Cracks Unsolved 1941 Enigma Message in Two Days
OpenAI GPT-6 Astra cracked an unbroken 1941 Enigma cipher on its own after humans failed for 21 years.
Research · Source: Hacker News
What happened
On September 15, 2026, researcher Carter Leffer submitted a successful break of German Army Enigma message Nr. 172, MVUEH, from July 10, 1941. The cipher had resisted all cryptanalysis attempts since 2005 on the Crypto Cellar Research project.
OpenAI GPT-6 Astra performed the break autonomously after Leffer pointed it at a list of unbroken ciphers. The model singled out message MVUEH, deduced that its plaintext resembled another broken message from the same day, and wrote custom Python and C++ software for an Enigma simulator and Bombe to search the keyspace.
The recovered key used wheel order 253, unlike the expected 512 order from other intercepts that day. The model successfully handled a rare turnover on the left-hand wheel at the 72nd letter and multiple original transcription errors, completing within two days what human cryptanalysts estimated would take weeks or months.
Key facts
- 2005 — Year since the MVUEH Enigma message resisted decryption attempts
- Nr. 172 — Log number of the unbroken MVUEH message from July 10, 1941
- 253 — Wheel order recovered by GPT-6 Astra, defying the expected 512 setup
- 72nd letter — Point where the Enigma left-hand wheel turned over, confounding earlier breaks
- 2 days — Time GPT-6 Astra took to complete research that takes humans months
Why it matters
This is not simple pattern recognition or code generation in a vacuum. The model autonomously combined historical archive research, hypothesis generation, custom software development, and iterative search to solve an open research problem.
It signals that modern frontier models can act as self-directed technical detectives. They do not just write snippets; they orchestrate multi-language tools, cross-reference external archival records like the German Bundesarchiv, and execute end-to-end attacks on complex data structures.
For builders
Automate domain-specific reverse engineering
Builders can now wrap agentic models around legacy protocols, proprietary file formats, and historical data dumps. Enterprises holding decades of untracked archives will pay for agents that write custom parsers to decode their backlog.
Synthetic tool-building replaces rigid pipelines
Instead of hardcoding processing pipelines, agents can author and compile high-performance C++ and Python tooling on the fly. Teams building rigid ETL engines lose out to dynamic agentic solvers that pivot when data formats mutate.
My take
Stop using models just to summarize emails and auto-complete React components. If a model can spin up a custom C++ Bombe and crack an eighty-year-old military cipher by digging through German archive records, your entire definition of an agent workflow is set way too low.
Original reporting: Hacker News. This is my rewrite and opinion.