OpenAI's GPT-6 Astra Deciphers WWII Enigma Message in 10 Hours

By AI123

In September 2026, OpenAI's GPT-6 Astra model deciphered a Second World War Enigma machine message in about 10 hours. The encrypted radio message, sent by a German soldier on July 10, 1941, contained only 82 characters and gave route directions related to the town of Rozenau. It had remained unresolved in the CryptoCellar public archive for more than 80 years.

Carter Leffin, who works in product development at Bloomberg in New York, asked GPT-6 Astra to attack the ciphertext. The model compared the message with previously decoded messages from the same day and found that “ROSENOW” was likely present in the unsolved text. It then built its own Enigma simulator, wrote attack code, and tested multiple key combinations against the message header. According to the report, the model tried about 14.8 million keys before recovering the plaintext. The German text it produced kept the sender's spelling error “BTTE” instead of “BITTE”, a detail that adds credibility to the result.

The report also notes that the Enigma was a hardware cipher from the 1930s and becomes far less secure once some context is known; a typical laptop is enough to handle it. Modern encryption is fundamentally different in strength. Using Bitcoin's 256-bit key based on the secp256k1 curve as an example, breaking it would require roughly 2^128 operations, a scale beyond the number of visible atoms in the sky. The exercise therefore does not threaten the mathematical basis of such systems. The report adds that real risks are more likely to come from implementation flaws, such as random number generator defects or poor key management, rather than brute-force attacks on the algorithm itself.

Explore topicsGPTModelsOpenAI

Source

View all