Amateur hobbyists crack historic ciphers with AI model GPT-6 Astra

Carter Leffen, a product-development coach at Bloomberg Law, prompted OpenAI’s GPT-6 Astra to tackle an unsolved German Enigma message dated 10 July 1941. The AI generated code to emulate the Enigma machine, explored over four billion rotor configurations, and produced a readable plaintext, which cryptographer Frode Weierud later verified as using distinct settings from a contemporaneous solved message. The breakthrough sparked excitement among cryptography specialists, who note a surge of hobbyists employing AI for similar tasks but also face a flood of inaccurate claims.

Following Leffen’s success, SentinelOne staff engineer Carter Church fed a scanned 1809 Italian cipher to GPT-6 Astra; the model extracted the symbols, located a partial key on a French cryptography site, and applied simulated annealing to uncover a key within six hours, revealing a report on troop positions that was sent to hobbyist Satoshi Tomokiyo for confirmation. Both episodes demonstrate that large language models can dramatically shorten the time needed to break historical codes, though professional validation remains essential.