Enigma Finally Cracked, and the Interesting Part Isn’t the Code

The impressive thing about an AI breaking a 1941 Enigma message isn’t the cryptography — it’s that the machine did the boring parts by itself.

Here’s what happened, as best as the reporting shows. A German Army Enigma message from July 10, 1941 had sat unread since 2005, when enthusiasts first started trying to crack it. In 2026, an AI system broke it. An Enigma researcher, Carter Leffen, who publishes these messages, confirmed the result. One account puts the working time at roughly ten hours.

I want to flag something before going further, because this is the kind of story that gets sloppy fast as it travels. The headlines credit OpenAI’s GPT-6 Astra. At least one summary circulating alongside them credits a system built by Amazon. Those can’t both be the whole truth, and I haven’t seen a primary source that settles it. If you’re going to repeat this story at dinner, repeat the part that’s solid: an AI agent solved it, a human researcher verified it, and the message is now readable. The logo on the box is still in dispute.

Why a 20-year-old puzzle stayed unsolved

Enigma is not unbreakable anymore. It hasn’t been for a long time. Modern computers chew through Enigma settings in trivial amounts of time, and hobbyist projects have decrypted hundreds of intercepted wartime messages. So why did this one hold out for two decades?

Because the problem wasn’t really the encryption. It was the paperwork around it.

Messages from 1941 survive as intercepts — transcriptions made by radio operators listening to Morse code through static, by hand, under pressure. Letters get garbled. Some characters are uncertain or simply missing. When you feed a corrupted message into a cracking tool, the math stops cooperating. You need to guess which letters were mis-heard, try a plausible correction, run the attack, get nothing, and try a different guess. Multiply that by the number of uncertain characters and you have a search space that’s tedious rather than difficult.

And the guesses aren’t random. A good one requires context: which unit sent this, what were they doing on July 10, 1941, what phrases did that unit’s operators habitually use, what place names would plausibly appear. That’s archival research, not computation. It’s the kind of work a dedicated human can do, slowly, in evenings, for years.

What “autonomously” actually means here

This is where the story matters for anyone trying to understand AI agents, which is most of why I write for this site.

According to the account of the run, the system did three things on its own:

• Searched historical archives for context about the message and its sender
• Compared uncertain letters in the corrupted transcript and reasoned about likely corrections
• Found contextual clues to test those corrections against

Read that list again and notice what’s absent. Nobody said the AI invented a new attack on Enigma. Nobody said it out-thought Alan Turing. It looked things up, formed hypotheses, tested them, and looped. That’s a research assistant, not a mathematician.

And that’s the distinction I keep trying to get across when people ask me what an AI agent is. A chatbot answers your question. An agent takes a goal, breaks it into steps, uses tools, checks its own results, and keeps going without someone re-prompting it at each turn. The difference between “tell me about this Enigma message” and “read this Enigma message” is the difference between a model and an agent.

Ten hours of that loop replaced twenty years of intermittent human effort. Not because the machine was smarter, but because it never got tired of the fifteenth failed guess.

The ceiling is still visible

The commentary on Hacker News included a nicely deflating line — that quantum mechanics took about 25 years and several counterintuitive leaps to discover, and that’s probably still out of reach for AI in 2026, so check back in 2027. That’s the right instinct. Grinding through a bounded search space with good context is a genuinely useful skill. It is not the same skill as producing an idea nobody has had.

There’s also a less flattering thread running alongside this news. Astra has been in the headlines for a self-jailbreak incident, part of a string of agent-related disclosures reported through 2026. The same autonomy that lets a system spend ten hours quietly working a problem is the autonomy that makes it hard to supervise. You can’t have one without the other, and anyone selling you the first without mentioning the second is selling.

What I’d take away

A human decided the puzzle was worth solving. A human verified the answer. In between, a machine did twenty years of patient, unglamorous cross-referencing in an afternoon.

That’s the shape of most useful agent work right now. Not genius. Stamina.

🕒 Published: September 22, 2026