GPT-6 Astra helped a researcher recover the key for MVUEH, a difficult 1941 German Army Enigma message, by finding a useful crib, writing search software and orchestrating a large constrained computation. The result matters as a concrete example of long-horizon tool use, but the evidence does not support headlines claiming that a model autonomously solved naval Enigma by intuition.
Key facts
- The target was German Army message MVUEH, received 10 July 1941, not naval M4 traffic.
- The documented search covered roughly 4.29 billion combinations under stated assumptions.
- The result uses a 14-letter crib from a neighboring solved message.
- Primary source: the MVUEH case study.
The archival distinction is central. CryptoCellar's record identifies MVUEH as a three-wheel Army Enigma message; the well-known M4 project was a separate attack on four-wheel naval messages. The project says researcher Carter Leffen set the target while Astra and specialist agents examined the archive, generated Python and C++ Enigma/Bombe-style programs, and ran searches. A neighboring message, SIPVX, supplied ROSENOW ROSENOW as a known plaintext fragment.
That fragment is like finding a few letters of a crossword answer in the same hand as the unsolved clue. It does not answer the puzzle directly, but it turns an unconstrained search into a structured test. The case study reports a search over rotor behavior, crib placement and plugboard settings, followed by millions of physical-key checks. It recovered reflector B, rotors II-V-III, rings HMF, body start RWD and ten plugboard pairs. The resulting plaintext is coherent German military traffic requesting a route of march and an immediate radio reply.
The result's best evidence is reproducibility rather than grandeur. The case study says independent implementations reproduce both message body and header in both directions, and the archive explanation describes why the short, error-prone text and unusual rotor turnover had resisted standard work. “The researcher set the goal and pushed the investigation forward,” the case study says—an unusually clear limitation that should survive retelling.
The strongest counterargument is right: crib attacks, rotor simulators and key searches are established cryptanalysis, and the system had a curated corpus, a prior neighboring break, human direction and a validation framework. There is no published OpenAI research claim independently validating the result; OpenAI's Astra launch does not mention MVUEH. What the project does demonstrate is substantial composition: agents can turn broad research objectives into source work, code, computation and checked outputs. That makes it relevant to research automation, while leaving the core historical and mathematical method firmly in human-understood territory.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)