Two cryptanalysts have used frontier language models to solve separate Enigma messages that had resisted previous efforts. OpenAI’s GPT-6 Astra decoded one message after searching archives, finding contextual clues and building an Enigma simulator, while Anthropic’s Claude Opus 5 solved another with more extensive guidance from a human researcher.
Developer Carter Leffen asked Astra to find and decode an unsolved item in a public message database. Retired engineer Frode Weierud, who maintains the Crypto Cellar archive, validated the recovered plaintext and said the work compressed weeks or months of research into two days. In the second case, cybersecurity executive Jack Willis helped Claude exploit the known signature of a German officer’s name to reach a solution. Seven Enigma messages remain unbroken, according to Weierud.
The achievement combines tool use, coding and archival research rather than proving a general breakthrough in cryptography. Astra’s logs also referred to messages in a private collection, and Weierud could not determine whether the model had accessed those files or found copies in public German archives. That uncertainty makes provenance an important part of evaluating agentic research, even when a human expert confirms the final answer.