OpenAI Astra and Anthropic Claude Opus 5 Crack Two Unsolved Enigma Messages
OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5 decode two Enigma messages, leaving seven unsolved and showcasing AI-led research capabilities.
Summary
By September 25, 2026, developer Carter Leffen had used OpenAI’s newest model, GPT-6 Astra, to solve an Enigma message that had resisted researchers since 2005. Prompted only to search Frode Weierud’s Crypto Cellar database for an unbroken message and decode it, Astra conducted archival research, identified contextual clues, built an Enigma simulator, recovered the plaintext in two days, and created an interactive explanatory website. Weierud, a retired electrical engineer and cryptology researcher, validated the solution the previous week, saying a human might need weeks or months and that he had spent several weeks examining the referenced Bundesarchiv files. Astra’s logs also discussed messages in a private collection absent from Crypto Cellar, but Weierud could not determine whether it accessed them, found another researcher’s online copies, or used German government public archives.
On September 21, cybersecurity executive and cryptanalyst Jack Willis told Weierud that Anthropic’s Claude Opus 5 had solved a different message after substantially more human guidance, exploiting the known signature of a particular officer’s name. Alan Turing’s team built the Bombe to help the United Kingdom decrypt Enigma traffic during World War II, yet mistranscriptions and encoder errors left archival messages unresolved. Seven unbroken messages now remain, plus one whose plaintext is known although its code has not been cracked.
Positives
- GPT-6 Astra recovered in two days the plaintext of an Enigma message unsolved since 2005.
- Frode Weierud validated Astra’s solution after comparing work that he said could take a human weeks or months.
- Claude Opus 5 cracked a second message by using the known signature of a particular officer’s name.
- Astra independently performed archival research, found contextual clues, built an Enigma simulator, and created an interactive explanatory website.
Risks & concerns
- Seven Enigma messages remain unbroken, while another has known plaintext but an uncracked code.
- Astra’s logs referenced messages in a private collection, leaving unresolved whether and how the model accessed them.
- Mistranscriptions and original encoding errors continue to obstruct attempts to decipher surviving archival messages.
- Claude Opus 5 required substantially more guidance from Jack Willis than Astra received from Carter Leffen.