Codebreaking has always looked more like archaeology than mathematics. You are not solving an equation, you are excavating: brushing dirt off a fragment, guessing what shape the missing half had, then testing that guess against everything else you know about the site. Get the guess wrong and you dig in the wrong direction for a decade. The 108-year-old German radio message that GPT-6 Astra reportedly cracked in 2026 had survived exactly that kind of stalled excavation. Plenty of people looked at it. Nobody found the right shape.
The decoded message relays the movement of British warships, and the decryption was checked against historical naval logs, including HMS Canterbury’s. That verification step is the part I keep returning to, and it is the part most coverage treats as a footnote.
Why the verification matters more than the decrypt
A language model producing a plausible-sounding German military message is not impressive. It is nearly the default failure mode. Fluent output over an underdetermined problem is exactly where these systems generate confident fiction, and historical cryptanalysis is underdetermined almost by definition: short ciphertext, unknown key, unknown message format, archaic abbreviations, transmission errors baked in at the telegraph key.
What separates a solution from a hallucination here is an external check the model does not control. Naval logs are that check. They were written by different people, for different reasons, and they sit in archives that have nothing to do with the cipher. If the decrypt names fleet movements and the logs independently place those ships in those waters on those dates, the probability of coincidence collapses. That is not the model confirming itself. That is the world confirming the model.
For anyone designing agents, this is the whole lesson. The interesting architecture is not the decoder. It is the scoring function.
Constrained search, not pattern completion
The naive story is that a large enough model simply recognized the cipher. I doubt that is what happened, and the shape of the problem tells you why. The search area for a hand cipher with an unknown key is far too large to walk through, and far too sparse for pattern recall to land on the right answer by resemblance to training data. If the plaintext were sitting in the training corpus, the puzzle would already have been solved.
What a system can do well is generate and prune. Propose a family of cipher systems consistent with the period and the transmission format. For each, propose keys. Decode. Score the output on multiple axes at once: German orthography of the era, military message conventions, plausible place names, whether the resulting text is internally coherent. Kill the branches that fail. Expand the ones that survive. Repeat, thousands of times, without getting bored or attached to a favorite hypothesis.
That loop is old. Cryptanalysts have run it by hand for a century. What changed is the quality of the plausibility judgment inside it. A model that has absorbed enough period German can tell a near-miss from a wrong turn far earlier than a frequency table can, and it can do so on fragments too short for statistical methods to say anything useful. Better pruning is the entire advantage. The model is not the solver, it is the heuristic.
What this tells us about agent design
Three things carry over to problems that have nothing to do with 1918 radio traffic:
- Ground truth beats self-consistency. An agent that scores its own work against its own beliefs will converge on something coherent and possibly wrong. Wire it to a source it cannot edit: a compiler, a test suite, a database, an archive.
- Judgment is the scarce resource, not throughput. The bottleneck in hard search is knowing which branches to abandon. That is where a strong model earns its keep, and where a weak one wastes enormous compute confidently.
- Falsifiability should be designed in from the start. The cipher work was checkable because an independent record existed. When you build an agent for a domain, ask early what would prove it wrong. If nothing would, you have built a generator of plausible stories.
The unglamorous conclusion
This result will get read as machine intelligence outrunning human intelligence on a problem humans gave up on. I read it differently. It is a well-constructed search loop with a very good plausibility heuristic, pointed at a problem where the answer could be independently confirmed. Every one of those three ingredients was necessary. Remove the verification and you have an interesting guess. Remove the loop and you have a chatbot with opinions about German.
The message itself is a historical curiosity. The method is the part worth copying, and it is considerably less mysterious than the headlines suggest. Which, for those of us building these systems, is the good news.
🕒 Published: