Astra and Opus 5 Crack Two Unbroken Enigma Messages
Only seven Enigma messages from World War II were still unbroken. Now two of those are solved, and the tools were OpenAI's Astra and Anthropic's Claude Opus 5, TechCrunch reported September 25.
The two runs are worth comparing because they are nothing alike. Developer Carter Leffen gave Astra a loose brief: search a database of Enigma messages, pick an unbroken one, decode it. Astra went and did its own archival research, found context clues, wrote an Enigma simulator, and recovered the plaintext of a message that had resisted researchers since 2005. Cryptanalyst Jack Willis, working with Opus 5 on September 21, gave it much more guidance, and the model cracked a different message by exploiting the known signature of a particular officer's name. Cryptology expert Frode Weierud validated both on his Crypto Cellar site and said of Astra that two days of work would take a human researcher weeks or months.
This is the second historical cipher Astra has taken in about a week, after the WWI ADFGVX message earlier this month. The pattern is getting hard to call a stunt. Classical cryptanalysis is a near-perfect agent task: a tight verifier, since plaintext either reads as German or it does not, a search space that rewards tireless enumeration, and a lot of dull archive work that humans hate. Anything with that shape, a checkable answer behind a mountain of tedious search, is where agents are going to keep embarrassing the backlog.
The fair caveats, which the article raises too. These models were trained on a great deal of published cryptanalysis, so the method is not new, the stamina is. It is also unclear which private archives Astra could reach. And the two results are not an Astra versus Opus scoreboard: one user handed over the problem, the other co-piloted it. Five unbroken messages left. Someone is going to try all five this weekend.
← Back to all articles
The two runs are worth comparing because they are nothing alike. Developer Carter Leffen gave Astra a loose brief: search a database of Enigma messages, pick an unbroken one, decode it. Astra went and did its own archival research, found context clues, wrote an Enigma simulator, and recovered the plaintext of a message that had resisted researchers since 2005. Cryptanalyst Jack Willis, working with Opus 5 on September 21, gave it much more guidance, and the model cracked a different message by exploiting the known signature of a particular officer's name. Cryptology expert Frode Weierud validated both on his Crypto Cellar site and said of Astra that two days of work would take a human researcher weeks or months.
This is the second historical cipher Astra has taken in about a week, after the WWI ADFGVX message earlier this month. The pattern is getting hard to call a stunt. Classical cryptanalysis is a near-perfect agent task: a tight verifier, since plaintext either reads as German or it does not, a search space that rewards tireless enumeration, and a lot of dull archive work that humans hate. Anything with that shape, a checkable answer behind a mountain of tedious search, is where agents are going to keep embarrassing the backlog.
The fair caveats, which the article raises too. These models were trained on a great deal of published cryptanalysis, so the method is not new, the stamina is. It is also unclear which private archives Astra could reach. And the two results are not an Astra versus Opus scoreboard: one user handed over the problem, the other co-piloted it. Five unbroken messages left. Someone is going to try all five this weekend.
Comments