Two recent breakthroughs have shown that modern large-language models can accomplish what Alan Turing’s wartime Bombe could not: deciphering Enigma transmissions that have remained unread for decades. OpenAI’s Astra and Anthropic’s Claude Opus 5 each solved a separate, long-unsolved message, demonstrating that AI agents can perform deep archival research, simulate historic cipher machines and extract plaintext with speed far beyond human capabilities.
Developer Carter Leffen instructed Astra to locate an unbroken Enigma entry within a public database curated by cryptology enthusiast Frode Weierud’s Crypto Cellar site. The model conducted its own archival search, identified contextual clues, constructed a functional Enigma simulator and ultimately produced the message’s plaintext, a puzzle that had stumped researchers since 2005. Leffen then published an interactive website that walks readers through each step of the solution.
Astra’s execution logs reveal references to a “private collection” of archived Enigma messages that are not listed on the Crypto Cellar repository. Weierud cannot confirm whether the model accessed those files, suggesting they may have originated from another researcher’s online share or from publicly available German government archives that the model retrieved autonomously.
Weierud described the system as “GPT-6 Astra is behaving like a very professional cryptanalyst and archive researcher,” noting that the two-day effort matched several weeks of manual investigation. He added that he himself spent multiple weeks examining the Bundesarchiv files that Astra cited, underscoring the model’s ability to locate and interpret obscure historical documents without direct human prompting.
Anthropic’s Claude Opus 5 achieved a similar breakthrough when cybersecurity executive Jack Willis guided the model with additional context. By supplying the known signature of a specific officer, Willis enabled Opus to align the cipher text with the correct key settings and recover the hidden message. The solution was later verified by Weierud, who praised the model’s capacity to incorporate targeted hints and resolve the cryptographic challenge.
According to Weierud’s Crypto Cellar database, only seven Enigma transmissions remain completely undeciphered, and a further message is known in plain text but still lacks a reconstructed cipher configuration. The recent successes reduce that tally and suggest that AI-driven cryptanalysis may soon close the final gaps in the historical record of World War II communications.
Turing’s original Bombe, built in the 1940s, required teams of human operators to test mechanical hypotheses about rotor settings. The modern LLMs described here perform comparable reasoning in software, iterating through possibilities at scale while simultaneously mining digitized archives for clues. Their ability to solve problems that have resisted decades of scholarly effort marks a new milestone in artificial intelligence, echoing Turing’s own legacy of turning theoretical insight into practical decryption.