GPT-6 Astra solves unsolved WWI German radio cipher

GPT-6 Astra, a recent model release, has successfully decoded a World War I German radio message that has remained unsolved for over a century. The message, transmitted on November 27, 1918, was part of a set of ADFGVX-encoded communications listed on Scienceblogs.de, a German science blogging portal that maintains a list of 50 unsolved ciphers. According to the detailed analysis published on PrinzAI, the model identified the encryption key as “TRUPPENVERSCHIEBUNG” and reconstructed the original plaintext, revealing information about the arrival of the English cruiser HMS Canterbury in Sevastopol.

This achievement is notable because the ADFGVX cipher, a combination of a substitution and a transposition cipher, was considered unbreakable during the war and was not fully understood until the 1980s. While hundreds of ADFGVX messages have been decoded in the decades since, including by codebreaking expert George Lasry, this specific message had eluded efforts to solve it. The model’s success demonstrates a significant leap in the ability of large language models to perform complex, multi-step logical reasoning and pattern recognition tasks that previously required specialized human expertise.

The details of the decryption

The ADFGVX method uses a 6x6 grid where each cell is identified by a pair of letters from the set {A, D, F, G, V, X}. The encryption process involves two main steps: substitution and transposition. First, the plaintext is converted into pairs of letters using a key-based table. Second, these pairs are rearranged using a transposition key. In this case, the model used “TRUPPENVERSCHIEBUNG” as the encryption word, a term referenced in J. Rives Childs's “The History and Principles of German Military Ciphers, 1914–1918.”

The decryption process described in the source is intricate. The key “TRUPPENVERSCHIEBUNG” is first rearranged alphabetically to determine the column order. The encrypted message, consisting of 170 characters, is then written into rows of 19 symbols. This results in 8 full rows and one partial row of 18 symbols. The model calculated the position of each letter in the transposed message by counting symbols based on the alphabetical order of the key. For example, the letter “T” is the 16th letter in the alphabet, so the model counted 14 full columns of 9 symbols plus one column of 8 symbols to find the 135th symbol, which was “A.” This process was repeated for each letter to reconstruct the original plaintext.

The model’s hypothesis for why this message was previously unsolved is that the key “TRUPPENVERSCHIEBUNG” was officially used starting on December 9, 1918, which is after the date of transmission (November 27, 1918). This discrepancy suggests that the key may have been used earlier than historical records indicate, or that there was an error in the historical documentation. The model verified its work by cross-referencing the decoded message with historical logs, confirming that HMS Canterbury arrived in Sevastopol on November 24, 1918, and that an allied squadron arrived on November 26, 1918.

Context and significance

The ADFGVX cipher was a significant advancement in cryptography during World War I, combining the security of a substitution cipher with the complexity of a transposition cipher. It was used by the German military to secure their communications and was considered unbreakable by Allied codebreakers at the time. The cipher was not fully understood until the 1980s, when researchers began to decode historical messages. The fact that GPT-6 Astra was able to solve a message that had resisted human efforts for over a century highlights the potential of AI to tackle complex problem-solving tasks that require both logical reasoning and pattern recognition.

This result is part of a broader trend of AI models being used for tasks that were previously considered exclusive to human experts. Large language models have shown impressive capabilities in areas such as mathematics, programming, and scientific reasoning. The ability to solve a complex cryptographic puzzle like the ADFGVX cipher suggests that these models are becoming increasingly capable of handling tasks that require multi-step reasoning and the integration of multiple pieces of information. This has implications for fields such as cybersecurity, where the ability to break or analyze ciphers is crucial.

What it means for developers

For developers and researchers, this result demonstrates the potential of AI models to be used for complex problem-solving tasks. The ability to solve a cipher like the ADFGVX requires not only pattern recognition but also the ability to follow a multi-step logical process. This suggests that AI models can be used for tasks that require both creativity and precision, such as debugging complex code, optimizing algorithms, or analyzing large datasets. Developers should consider how these models can be integrated into their workflows to enhance productivity and solve problems that were previously too complex or time-consuming to tackle manually.

However, it is important to note that AI models are not infallible. The model’s hypothesis about the date of the key’s usage contradicts historical records, suggesting that there may be errors in the historical documentation or that the model made an assumption that was not fully justified. Developers should be cautious when relying on AI models for critical tasks and should always verify the results using independent methods. This case serves as a reminder that while AI models are powerful tools, they are not a substitute for human judgment and expertise.

What to watch

• Historical verification: Researchers should verify the model’s hypothesis about the date of the key’s usage by cross-referencing with other historical sources. This could provide new insights into the history of cryptography during World War I.
• Model capabilities: The ability of GPT-6 Astra to solve this cipher suggests that newer models may be even more capable of tackling complex problem-solving tasks. Developers should keep an eye on future model releases and their performance on similar tasks.
• AI in cybersecurity: The ability of AI models to break or analyze ciphers has significant implications for cybersecurity. Researchers should explore how these models can be used to improve security measures and protect against cyber threats.
• Ethical considerations: The use of AI models for tasks such as cryptography raises ethical questions about privacy and security. Developers and policymakers should consider the implications of AI’s growing capabilities in this area and develop guidelines for its responsible use.