GPT-6 Astra Breaks an Old Enigma Message - Schneier on Security
Short summary
GPT-6 Astra autonomously broke a previously unsolved Enigma message from WWII by developing its own cryptanalysis software and identifying a repeated phrase as a crib for attack. The breakthrough sparked similar successes, with Claude Opus 5 subsequently breaking another long-encrypted message using the same approach.
- GPT-6 Astra independently selected message Nr. 172 (MVUEH) as the most promising target, then built Python and C++ Enigma simulators and a Bombe implementation to crack it
- The AI identified "ROSENOW ROSENOW" as a repeated plaintext fragment and used it as a known-plaintext attack vector to recover the key
- Claude Opus 5 followed up by breaking message Nr. 205 using a signature phrase (XHARTJENSTEINX) as the crib, showing the technique's reproducibility
Full summary
GPT-6 Astra cracked an Enigma message that had remained unbroken for decades—message Nr. 172 (MVUEH)—without human guidance beyond the initial request to attempt breaking unbroken messages from the Crypto Cellar Research website. The AI analyzed the available ciphertexts, identified MVUEH as the most promising target, and suspected it was related to another message (Nr. 173, SIPVX). It then built Python and C++ software to simulate an Enigma machine and implement a Bombe (the electromechanical device used historically to crack Enigma). Using the repeated place name ROSENOW ROSENOW as a known plaintext crib, it systematically worked through possibilities until finding the correct key and plaintext. The whole process happened autonomously—the AI didn't just execute instructions but made strategic decisions about which message to attack and what approach to take.
The breakthrough immediately inspired similar attacks. On September 20, 2026, Anthropic's Claude Opus 5 successfully broke another long-standing message (Nr. 205/Nr. 285, FMNGI) from July 1941. This attack also relied on a crib, specifically the frequently occurring signature XHARTJENSTEINX. Both breaks demonstrate that modern large language models can handle the kind of cryptanalytic reasoning and software development that historically required human expertise.
The implications are worth taking seriously. These weren't toy problems—these were real Enigma messages that had withstood decades of attempts. That AI systems can now independently identify attack strategies, write the necessary cryptographic software, and execute complex breaks suggests a genuine shift in what's computationally feasible. The fact that two different AI systems (GPT-6 and Claude Opus 5) independently succeeded within days of each other suggests this isn't a fluke but a real capability.