OpenAI GPT–6 Astra breaks Enigma message that has resisted solution since 2005

TL;DR Highlight

GPT–6 Astra broke the Enigma message MVUEH, unsolved since 2005, in two days

Who Should Read

Backend developers curious how LLM agents write code and run tools to solve long-running problems

Core Mechanics

The MVUEH message was broken
- The 10 July 1941 Enigma message MVUEH (Nr. 172) had resisted solution since 2005.
- On 15 September 2026, Carter Leffen sent his break to the author for validation.
The key differed from the daily key
- Other messages that day used wheel order 512, but MVUEH used 253.
- Transcription errors and a left-wheel turnover at the 72nd letter also made it harder to break.
Astra picked the target itself
- Leffen only asked it to try breaking any unbroken messages on the site.
- Astra chose Nr. 172 as most promising and suspected its plaintext related to Nr. 173 (SIPVX).
ROSENOW ROSENOW was the crib
- A crib is a word you guess appears in the plaintext.
- Astra used it, wrote Python and C++ code for an Enigma simulator and Bombe, and found the key.
Logs show archive research

Community Reactions

Opinion that "entirely on its own" is overstated
- Commenters noted it wrote Python and C++ simulators, which sits oddly with that claim.
- They asked how novel that software was and how much of the break it did.
Opinion stressing the human role
- One commenter said the two-day collaboration shows the importance of a human in the loop.
- Another complained that articles ignore the person driving such breakthroughs.
Experience reproducing it with another model
- One commenter said Gemini 3.8 flash decrypted the ciphertext in about 45 minutes in an unsteered run.
- Opus was still running a roughly 100-minute script.
Counterargument about open-source reuse
- Critics said Astra just used existing open-source Enigma simulators and Bombes and tried many cribs.
- Some argued training data helped it guess cribs.

How to Apply

Give an agent an open-ended goal
- Spend 30 minutes today handing an unsolved bug or log problem to an agent, including target selection.
- Keep the instruction simple: pick the most promising item and explain why.
Make the agent build its own verification tools
- Have it write a simulator or script to test hypotheses and keep run logs.
- Read the logs afterwards to check the assumptions it used.
Define the success check first
- State the done criterion in the prompt before starting, such as matching a re-encrypted result.
- Require the agent to verify and report whether it passes.

Try It

copy & run

# Paste into Claude or ChatGPT as is (simplified version of the article's method)

You are a cryptanalysis researcher. Solve the problem below.
Problem: {{description of the unsolved problem}}
Materials: {{logs, messages, data}}

Steps:
1. Pick the single most solvable item and explain why.
2. List 3 hypotheses about likely plaintext clues (the equivalent of cribs).
3. Write a short Python script to test the hypotheses.
4. Define the done criterion first, then check whether the result meets it.

Next: afterwards, ask it to separate the assumptions used from the failed attempts.

Terminology

EnigmaA WWII German cipher machine. Its wheel and plugboard settings act as the password.

cribA word you guess appears in the plaintext. It is like pre-filling part of a puzzle's answer.

BombeA machine that searches for Enigma settings. Here Astra built it in software.

TurnoverThe moment a wheel advances its neighbor by one step. It is rare and complicates breaking.