For over a century, a 170-symbol German ADFGVX radio message from November 27, 1918 sat unsolved in a collection of World War I cryptograms. Now, an experiment with OpenAI’s GPT-6 Astra has reportedly produced a German plaintext that aligns with surviving Royal Navy records around Sevastopol. The result is a breakthrough for historical cryptography—and a reminder that we still don’t fully understand how frontier models reason.

The fact that an AI model can crack a cipher that resisted human cryptanalysts for generations is significant on its own. But the real story here is the uncertainty: GPT-6 Astra produced the plaintext, yet the process by which it arrived at the solution remains opaque. That’s both the promise and the problem with frontier AI.

What Happened

The ciphertext belongs to a collection of German ADFGVX radio messages from World War I, previously featured in ScienceBlogs’ Klausis Krypto Kolumne as part of its “Top 50 unsolved encrypted messages” series. The ADFGVX cipher was a German field cipher used late in the war, and surviving messages have remained a target for amateur and professional cryptographers alike.

According to the report, GPT-6 Astra produced a German plaintext that reportedly matches surviving Royal Navy records detailing naval movements around Sevastopol. This is notable because it suggests the decoded message isn’t just a plausible-sounding string of German—it actually corresponds to historical events documented independently by British naval forces.

The proposed reading points to real naval movements, lending credibility to the result. However, the report stops short of explaining how GPT-6 Astra reached its solution. Did it recognize patterns from training data? Did it apply a novel inference method? The article leaves these questions open, which is exactly why this story matters beyond the history books.

What’s most intriguing is what the result implies: if a model can crack a 108-year-old military cipher without explicit programming for that cipher’s structure, it suggests that modern LLMs are doing something far more sophisticated than pattern matching. That has implications for cryptography, historical research, and our understanding of how these models reason.

Read the full announcement →

My Take

Let’s be honest: a single, unverified decryption of a century-old cipher is not going to rewrite cryptography textbooks. But it does tell us something important about where AI capability is heading. GPT-6 Astra didn’t just brute-force letters—it apparently reconstructed meaning from a 170-symbol sequence with no known solution, producing text that aligns with historical records. If true, that’s not pattern matching; that’s reasoning.

The bigger issue is transparency. The fact that the report can’t explain how GPT-6 Astra solved the cipher is a recurring problem with frontier models. We get impressive outputs, but the reasoning remains a black box. For developers, this is a feature and a bug. It’s a feature when we want creative solutions to hard problems. It’s a bug when we need to validate the result in high-stakes contexts—like, say, verifying a military message from 1918.

The cryptography angle is also worth watching. If models like GPT-6 Astra can crack historical ciphers with minimal guidance, what does that mean for modern encryption? The answer is probably “not much” for well-designed modern algorithms, but it raises questions about the longevity of security schemes that rely on obscurity or weak key derivation. The bar for what counts as “hard to crack” is moving.

What to Watch

  • Verification: Expect historians and cryptographers to independently validate the plaintext against other naval records. Confirmation would elevate this from curiosity to landmark.
  • Methodology: Watch for any follow-up details on how GPT-6 Astra reached its solution. If OpenAI or researchers can reverse-engineer the reasoning, it could inform how we design AI systems for cryptography.
  • Implications for modern security: If frontier models can infer structure from sparse, noisy data at this level, expect renewed pressure on legacy encryption and a push toward post-quantum standards sooner rather than later.