Astra and Opus are tackling Turing’s other test

Introduction

Alan Turing is most closely associated with the test that asks whether a machine’s behavior can be distinguished from a human’s. During his lifetime, however, a more immediate challenge was breaking Nazi Germany’s Enigma communications during World War II. That historical effort was never quite finished: a small number of archived messages remain unresolved, often because of transcription mistakes or errors made by the original operators.

Two recent cases suggest that today’s frontier models may be useful in completing parts of that unfinished work. Cryptanalysts report that OpenAI’s Astra and Anthropic’s Claude Opus helped recover the plaintext of two separate Enigma messages that had resisted earlier attempts.

Key points

• Developer Carter Leffen asked Astra to search a database of Enigma messages for an unresolved record and attempt a decryption.
• Astra reportedly searched archival material, identified contextual clues, and built an Enigma-machine simulator as part of the process.
• The model recovered a message that had puzzled researchers since 2005. Frode Weierud, who maintains the Crypto Cellar resources, later validated Leffen’s solution.
• Astra’s logs referred to material in a private collection. Weierud could not establish whether the model accessed that collection, found related information elsewhere online, or drew on public German archives.
• Jack Willis later used Claude Opus to solve a different message, giving the model more direct guidance, including the known name of a specific officer.
• Weierud says seven unsolved Enigma messages remain, along with one case in which the plaintext is known but the cryptographic settings have not been recovered.

Why it matters

The significance of these cases is not simply that an AI model performed a large number of cryptographic guesses. Enigma research is also a problem of reconstruction. Researchers must connect imperfect ciphertext, historical records, operator habits, military context, and possible copying errors. Astra’s reported workflow is notable because it combined several kinds of work: finding sources, forming hypotheses, writing a simulator, and testing candidate solutions.

That combination resembles the growing class of AI agents that can pursue a multi-step research objective rather than answer a single question. Weierud wrote that what Astra accomplished in two days would have taken a human researcher weeks or months, although this is a personal comparison rather than a controlled benchmark.

The examples should not be treated as proof that models have replaced cryptanalysts. Claude Opus received a decisive human clue, and Astra’s precise information trail remains uncertain. A model may also rediscover a solution from scattered public material without making clear which source mattered. For historical and scientific work, reproducibility, source auditing, and independent review therefore remain essential.

Enigma offers a particularly demanding test of practical intelligence because the answer is not simply waiting in a clean dataset. The evidence is incomplete, the records can be wrong, and the model must often create the tools needed to examine its own hypotheses. As the remaining messages are investigated, the more useful question may be not whether AI “passed” a second Turing test, but how effectively people and models can divide the work—and how reliably the result can be checked.

Source: TechCrunch AI