Amateurs use GPT-6 Astra to crack 85-year-old German army code
Non-experts are using LLMs to break historical ciphers, leading to a surge in AI-generated solutions that professional cryptographers must now validate.
Amateurs are using the AI model GPT-6 Astra to crack historical codes, including a secret German army message that had been encrypted for 85 years [4].
The surge in non-expert cryptanalysis is flooding professional cryptographers with requests to validate AI-generated solutions [4]. This trend harnesses the ability of LLMs to rapidly deploy decryption algorithms and search the internet for extra historical documents or information [4].
AI-Powered Codebreaking Enters the Amateur Realm
The use of GPT-6 Astra to decode the 85-year-old German army message highlights a growing trend of non-experts applying LLMs to break historical ciphers [4]. Carter Leffen, a product-development coach at Bloomberg, successfully decoded the message despite having no prior experience in cryptography [4].
Leffen reported that he nudged the model several times and encouraged it to continue working overnight before arriving at a solution [4]. His case is part of a broader wave of amateurs who are now harnessing the ability of AI to rapidly execute decryption algorithms [4].
From the 'Crypto Cellar' to the Enigma Solution
To find the cipher, Leffen used the AI to search Crypto Cellar Research, a website of unsolved historical ciphers operated by independent cryptographer Frode Weierud in Oslo [4]. The model selected and then decoded a message encrypted by an Enigma machine [4].
The Enigma machine was used by the German armed forces during the Second World War to protect military communications [4]. These devices used a system of rotors and plugs to reroute electrical signals, creating up to 103 sextillion possible encryption combinations [4].
The Expert Backlash Against AI 'False Positives'
While amateurs are finding success, cryptography professionals report being overwhelmed by a deluge of requests to verify AI solutions [4]. Many of these submissions are incorrect, leading to a high volume of errors.
Elonka Dunin, a retired game developer and author on cryptography, noted that "There are so many false positives that are coming in from people who don’t really understand cryptanalysis" [4]. The lack of fundamental knowledge among users often results in AI-generated solutions that can be incorrect [4].
Pattern Recognition Beyond Cryptography
The application of LLM-style pattern recognition is extending into other scientific disciplines, including the study of Earth's prehistoric history [3]. Paleontologists are now using similar models to reconstruct biodiversity timelines by analyzing the rock record [3].
One AI model trained on nearly four million fossils suggests that three of the "big five" mass extinction events were gradual declines in biodiversity [3]. This contradicts earlier reconstructions that described these events as sudden cataclysms [3].
What we don't know
- How many of the historical ciphers decoded by LLMs have been officially validated as correct by professional cryptographers?
- What specific prompts or "nudges" were used by Leffen to steer GPT-6 Astra toward the Enigma solution?
- Whether these AI tools are identifying genuine patterns or hallucinating plausible-looking but incorrect decryptions in complex ciphers.
