Amateurs Use AI to Crack Long-Secret Historical Codes

Summary

Amateurs are increasingly using large language models to investigate historical ciphers, and cryptographers are receiving a surge of requests to validate their results. In one case, Bloomberg product development coach Carter Leffen asked OpenAI’s GPT-6 Astra to select and decode an unsolved 1941 German army message encrypted with an Enigma machine. The model used a previously solved message as a possible crib, wrote code to simulate Enigma, and tested more than four billion settings in ten hours. The resulting solution was sent to independent cryptographer Frode Weierud, who confirmed that the message used settings different from those of the related message, an unusual detail that may explain why it remained unsolved. In a second case, SentinelOne engineer Carter Church supplied GPT-6 Astra with an image of an 1809 cipher sent by an Italian official to a general in Napoleon’s army. The model transcribed the symbols, found a partial key in online records, and used simulated annealing to find a successful key in about six hours; the decoded message described troop locations and was sent for verification. Church said the cipher could have been solved without AI, but the model reduced the time and effort needed. Specialists welcome the new interest but warn that many proposed solutions are false positives from users who do not understand cryptanalysis.