Two cryptanalysts say OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus have decoded long-unsolved Enigma messages, offering a striking test of what AI agents can do when they research, simulate and reason rather than merely autocomplete.
Watch Desk analysis
What happened
TechCrunch reports that developer Carter Leffen asked Astra to search a database of Enigma messages, investigate the historical context, build an Enigma-machine simulator and recover the plaintext of a message that had resisted researchers since 2005. Leffen also used Astra to create an interactive explanation of the problem.
Frode Weirerud, a retired electrical engineer who runs the Crypto Cellar archive, validated Leffen’s solution and said he was “in awe”. Weirerud wrote that Astra behaved like “a very professional cryptanalyst and archive researcher”, while noting that work he had spent weeks doing was completed by the model in two days.
TechCrunch also reports that cryptanalyst Jack Willis used Anthropic’s Claude Opus 5 to decode a different unsolved message. Willis supplied more guidance, including the known signature of an officer’s name, so the two demonstrations are not equivalent tests.
Weirerud says seven Enigma messages remain unbroken, plus one whose plaintext is known although the code has not yet been cracked. The report does not establish that Astra accessed private archives, despite logs mentioning material from a “private collection”. Weirerud said he did not know how that reference appeared.
Why it matters
This is a more useful AI demonstration than another leaderboard victory. The systems were reportedly asked to investigate an obscure historical problem, assemble missing context, construct a working simulation and produce an answer that an experienced specialist could check. That is a compact picture of agentic research, with the human expert still very much in the loop.
It is also a reminder that a successful result does not automatically explain the route taken. Astra’s apparent references to archives it may not have accessed raise a straightforward question about provenance: did the model discover public material, encounter another researcher’s work or simply produce a plausible-looking trail?
Our read
The interesting breakthrough is not that an AI can perform one dramatic feat. It is that the task combined archival research, historical inference, code generation and cryptanalysis, then produced something a specialist could validate. That is the sort of workflow that could make AI genuinely useful to researchers.
But “cracked an Enigma message” is a headline, not a methods section. The evidence here comes from TechCrunch’s report and the named cryptanalysts, not an independently reproduced benchmark. Readers should admire the result and keep the lab notebook open. Even Turing would probably have preferred a reproducible method to a magic trick.
What to watch
- Whether the two plaintexts and the full decoding methods are published for independent reproduction.
- Whether the remaining unsolved Enigma messages yield to similar AI-assisted workflows.
- How often models invent or misattribute archival sources while doing historical research.
- Whether future systems can explain their reasoning without requiring an expert detective at the other end. Sources and evidence: TechCrunch’s 25 September report is the basis for the attributed claims, including the accounts from Carter Leffen, Frode Weirerud and Jack Willis. The supplied evidence does not independently establish Astra’s archive access or provide a reproduced technical evaluation.
Discussion spark: Should an AI-assisted historical discovery count as a breakthrough if experts can validate the answer but cannot yet reconstruct exactly how the model found it?
Sources and evidence
- Astra and Opus just passed Turing’s other test (25 September 2026, 17:24 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.