OpenAI's Astra and Anthropic's Claude Opus crack long-unsolved Enigma messages

Two Enigma messages that had resisted cryptanalysts for years have reportedly been solved with the help of large language models. According to TechCrunch, a developer working with OpenAI's Astra and a cryptanalyst working with Anthropic's Claude Opus 5 each recovered the plaintext of a message long listed as unbroken, shrinking the pool of unsolved Enigma traffic to a handful of items.

Alan Turing is best remembered for the thought experiment about machine intelligence that bears his name, but his wartime codebreaking is the relevant backdrop here. Turing and his team built the Bombe, an early computing device that allowed the United Kingdom to read Enigma traffic during the Second World War. A small number of archival messages were never broken, usually because of mistranscription or errors by the original German encoders — mistakes that destroy the footholds codebreakers need.

Astra solved a message open since 2005

Carter Leffen, a developer, gave OpenAI's newest model a single instruction: search a database of Enigma messages, find one that had never been solved, and decode it. According to TechCrunch, Astra did its own archival research, gathered context clues, built a software simulation of the Enigma machine, and recovered the plaintext of a message that had stumped researchers since 2005. Leffen then used the model to build an interactive website walking through the problem and its solution.

Frode Weirerud, a retired electrical engineer with a lifelong interest in cryptology who maintains the Crypto Cellar website and its database of Enigma messages, validated the result. He told TechCrunch the solve left him in "awe."

An unclear research trail

One detail has proven harder to pin down. The Astra run's logs reportedly include discussion of archived messages held in a "private collection" that Weirerud does not host. He says he still does not know whether the model actually accessed them, and offers two possibilities: another researcher may have shared the material somewhere online, or the model may have reached the German government's public archives.

His overall assessment was striking. Quoted by TechCrunch, he wrote that "GPT-6 Astra is behaving like a very professional cryptanalyst and archive researcher," and that what the model achieved in two days would take a human researcher weeks or even months — adding that he had personally spent several weeks researching the Bundesarchiv files the model referenced.

A second break, with more hand-holding

The second solve came from a different direction. On September 21, cryptanalyst Jack Willis contacted Weirerud to report that he had used Anthropic's Claude Opus 5 to break a different unsolved message. Per TechCrunch, Willis provided significantly more guidance to Claude than Leffen had given Astra, and the break ultimately came from exploiting the known signature of a particular officer's name.

The two solves are therefore not equivalent demonstrations of autonomy: Astra worked from a bare instruction, while Claude operated with substantial expert direction. Together, they cleared two long-standing entries from the list.

What remains

Weirerud counts seven unbroken Enigma messages remaining, plus one unusual case where the plaintext is known but the encryption itself has never been reconstructed. Given this month's results, TechCrunch suggests those may not remain unsolved for long.

Why it matters

The story is a capability milestone rather than a product launch, and it is worth reading at two levels. The first is the historical payoff: messages sealed since the 1940s, one of them under active investigation since 2005, are now readable and have been checked by an independent domain expert.

The second is what the work reveals about agentic AI. The Astra solve was not a single-shot puzzle answer. It chained archival research, historical inference, software engineering and cryptanalysis across a multi-day run, starting from one plain-language instruction. Weirerud's comparison — two days of model work against weeks or months of expert human effort, on files he himself studied — is the kind of labour compression that matters far beyond wartime ciphers.

The details also raise a transparency question. If an agent's logs reference material in a private collection and even the validating expert cannot determine how it was obtained, provenance in AI-assisted research becomes a problem in its own right.

The caveats are real: two data points, one of them heavily guided, validated by a single expert rather than a formal benchmark. But as an existence proof that frontier models can finish open-ended, decades-old research tasks with minimal supervision, it is one of the clearer signals yet.