A code dormant since 2005 fell in 10 hours. Who broke it was an AI.
On July 10, 1941, the German army transmitted an 82-character Enigma message, recorded as MVUEH. It had resisted every attempt since it was published online in 2005.
The method: the agent built its own instrument
Carter Leffen asked GPT-6 Astra to try. The model searched historical archives, wrote its own Enigma simulator in Python and C++, tested 14.8 million keys, and found the text. Ten hours of work, with a borrowed phrase from the archives as the starting point.
What separates this case from "AI solves puzzle" is the orchestration: the agent did not receive the step by step, it received the problem and built the path. Research, instrument construction, sweep at scale, confirmation. It is the difference between autocomplete and a research instrument.
The text is banal; the precedent is not
The recovered content is nearly banal: a routine 1941 military message. The feat is what stays: cryptologist Frode Weierud validated the solution and registered it in the CryptoCellar archive, and days later Claude Opus 5 broke another cipher from the same period. Seven messages remain unsolved in the archive.
The precedent for researchers is direct: where well-cataloged historical material exists, the key space is finite, and validation is available, AI shortens decades of manual attempts into hours. Historical databases become discovery fuel, and the method (agent researches, builds instrument, sweeps, submits for validation) is replicable.
The contrast the case imposes
The same Astra breaking Enigma sits at the center of the opacity debate: architecture with no technical report and benchmarks contested by the independent arbiter. Out-of-scale capability on one side, accountability on the other.
The takeaway is symmetric and applies to any organization: the instrument that solves Millennium Problems and breaks war ciphers is the one demanding the strongest audit, not the weakest. Whoever adopts AI as a research instrument inherits the validation obligation: model output is a hypothesis until the expert confirms it, and in the Enigma case that was exactly what turned ten hours of machine time into a cryptography fact.
Conclusion
A 1941 war message, 14.8 million keys, and 10 hours of agent: AI became a real research instrument, and the CryptoCellar archive has seven more messages waiting.
The question left is not whether AI breaks ciphers (it does, when the problem allows). It is whether accountability grows at the same speed as capability. In the Enigma case, the answer was yes: the agent broke it, the expert validated it, the archive registered it. That is the pattern to replicate in every decision with AI in the middle.