Astra and Claude Break Two Enigma Messages

An AI Enigma break has now been independently documented for two separate German Army messages: one solution used OpenAI’s GPT-6 Astra and another used Anthropic’s Claude Opus 5. The important result is not that a chatbot guessed wartime text, but that human cryptanalysts could validate each proposed plaintext against the cipher records and historical context.

Key takeaways

Two different researchers used two different AI systems on separate archival messages.
CryptoCellar maintainer Frode Weierud documented or validated the resulting breaks.
The work does not imply that modern encryption has become vulnerable to language models.

How each AI Enigma break was checked

CryptoCellar’s MVUEH case study says developer Carter Leffen contacted Weierud on 15 September seeking validation of a message dated 10 July 1941. Astra had helped search archival material, construct analysis tooling and recover a proposed plaintext for a message that had resisted solution since 2005.

A second researcher, Jack Willis, used Claude Opus 5 on another unsolved message with more guidance, including a suspected officer-name pattern. TechCrunch reported both cases and noted that the amount of human scaffolding differed. That difference matters: the result demonstrates tool-assisted research, not a single repeatable button labelled “decrypt.”

Why validation is the real milestone

Historical cipher work can fail for mundane reasons: a copied character may be wrong, an operator may have broken procedure, or the surviving message may lack the context needed to recognise a plausible answer. A language model can search across archives, write simulators and combine clues quickly, but it can also invent a convincing narrative. Cryptographic consistency and external records are the check against that failure mode.

An AI Enigma break is credible only when the proposed plaintext fits the machine constraints, surviving ciphertext and historical record; fluent output by itself is not evidence of decryption.

The same rule applies to agentic software in current operations. Our Google PageBreak analysis explains why machine findings need reproducible proof, while our Gemini controls report shows how access boundaries matter when agents use external systems.

What the result does not mean

Enigma is a historical electromechanical cipher whose design, procedures and large body of surviving traffic have been studied for decades. Modern encryption relies on different mathematics, key sizes and threat models. Solving two archival messages with model-assisted research therefore says little about breaking contemporary end-to-end encryption.

The more defensible conclusion is narrower and still useful: AI systems can compress the slow parts of historical research—search, code generation, cross-referencing and hypothesis testing—when specialists keep the acceptance test outside the model.

Reproducibility is the next useful test. Other researchers should be able to inspect the ciphertext, document the corrections applied to archival material, rerun the mechanical steps and reach the same plaintext without relying on a model’s private chain of reasoning. That turns an impressive demonstration into a durable research contribution.

Facts at a glance

Fact	Detail
Validated update	24 September 2026
Systems	GPT-6 Astra and Claude Opus 5
Target	Two separate historical German Army Enigma messages
Control	Human cryptanalytic and historical validation
Claim boundary	No implication for modern encryption