OpenAI's Astra and Anthropic's Opus 5 crack unsolved Enigma messages

[Sequence]

this is a brilliant story about AI passing what it calls Turing's other test. Not the famous imitation game, but the work Alan Turing is best known for in real life, breaking Nazi Enigma codes. Two frontier models have now cracked Enigma messages that had gone unsolved for years. It is the kind of story that makes you smile. It is a lovely way to mark how far these systems have come

The first came from developer Carter Leffen, who used OpenAI's newest model, Astra. It decoded a message that had remained unbroken since 2005, and it did far more than brute force. The model searched archives, researched the historical context, built its own Enigma machine simulator and then recovered the plaintext. Leffen has put together an interactive website showing how it was done

The second was solved by Jack Willis, a cybersecurity executive, using Anthropic's Opus 5. That model broke a separate unresolved message by using patterns in officers' signatures as a guide. Both solutions were checked by Frode Weierud, a retired electrical engineer who runs Crypto Cellar, a database of cryptology material. He said Astra behaved like a professional cryptanalyst and did in two days what would take human researchers weeks or months. That is high praise from an expert

There are some caveats. It is not completely clear whether Astra had access to private archive collections or German government databases along the way, which could have given it a head start. And there are still seven unsolved Enigma messages out there, so the job is not finished. Still, the approach is what impresses me, planning the research and building its own tools rather than just crunching numbers

Turing would probably have loved this. The machines he helped create are now solving the very puzzles he worked on at Bletchley Park. Does anyone think the remaining seven will fall soon? And does it worry anyone that AI is this good at breaking codes, even very old ones?

Astra and Opus just passed Turing's other test | TechCrunchtechcrunch.com

[Will81]

Two days of work replacing weeks or months is the real headline. Historians and archivists have huge backlogs of documents nobody has time to read. Tools like this could open up so much material. Old letters, diaries and records in difficult handwriting could all become searchable

[ZenithCanopy]

Enigma is a solved problem in principle, so let's not get carried away. The machine's weaknesses have been known since the 1940s. What makes these messages hard is usually missing settings, bad transcriptions or garbled text. The model still had to work through all that, which is impressive. It is just not the same as breaking modern encryption

[Deep Anvil]

Bletchley Park should put these on display. A wartime codebreaking exhibit ending with AI solving the last unsolved messages would be a perfect story for visitors

[IonStorm99]

The archive access question needs answering. If Astra found a record somewhere with the key settings, the story changes quite a bit

It would still be clever research. But it would be finding the answer rather than cracking it, and those are not the same achievement. Hopefully Leffen's website clears that up

[Brad]

Seven left to go. I would bet on at least a few of them falling before the end of the year now that people know the approach works. Hobbyists will be trying every model they can get their hands on. It will turn into a bit of a race. That is great for the history of cryptography, whichever model wins

[Starforge Daniel]

Building its own Enigma simulator is the bit that blows my mind. That is the kind of thing a human expert would do as step one of a serious attempt

The model worked out the method, not just the answer. That is a big difference

[Gateway Coder]

Weierud's reaction is what convinced me this is significant. He has spent years with these messages and knows how hard they are. If he says it behaved like a professional cryptanalyst, I believe him

[Jackson_23]

Using officers' signatures as a clue is such a human approach. Codebreakers at Bletchley used exactly those kinds of habits, known as cribs, to get a foothold. Standard phrases and predictable sign offs were the weak point of Enigma

Opus picking up on that shows real understanding of how the system was broken historically. It did not just throw compute at it. That is proper detective work

[ScarletWrench]

Security people should take note, even if Enigma itself is ancient. The worry is not old ciphers but badly designed modern ones

Plenty of small companies roll their own encryption and get it wrong. A model that can study a system and find its weak point is a threat to those

[HeartbreakKidCurtis18]

Nice to see both big labs in the same story for something positive rather than a lawsuit. Makes a change from the usual headlines