AI cracks unsolved WWII Enigma cables in 10 hours

If you think frontier AI is only good for rewriting emails or generating code snippets, codebreakers just got a very different demo.

In September 2026, the venerable cryptanalysis archive CryptoCellar quietly updated its 1941 message logs. Next to several German military cables that had gathered dust for over 80 years, new credits appeared marking them solved by AI. One 82-character Enigma radio intercept, in particular, was cracked by GPT-6 Astra in about 10 hours.

To be clear, nobody is claiming Enigma itself was cracked for the first time. The machine mechanics were solved decades ago. The real bottleneck here was historical: the exact daily keys, rotor orders, and plugboard wirings for these specific transmissions vanished 85 years ago. When an encrypted message is just 82 characters long, standard statistical attacks usually choke.

More detective work than brute force

What makes this breakthrough wild is how Astra actually solved it. The model did not just hammer random permutations.

Instead, it conducted a full historical and forensic investigation. Facing faded and degraded original transcripts, the system combed through German archives, compared conflicting transcriptions to identify misread characters, and hunted for context in adjacent unit dispatches.

Then it built its own tools. Astra wrote an Enigma software simulator from scratch, generated scripts to explore plausible rotor settings, and spun up parallel experiments to test hypotheses. Roughly 10 hours later, the plaintext dropped out.

The decrypted message itself was remarkably mundane. A German unit was simply reporting its current location, asking which road to advance along, and requesting an immediate radio reply.

The researchers noted this was not a magic push-button operation. A human operator selected the target, fed in foundational historical data, and nudged the investigation when needed. But the heavy lifting—cross-referencing dirty archives, writing the simulator, and running the cryptanalysis loop—was carried out by the AI.

Why it matters

We see this as a huge inflection point for agentic workflows in real-world forensics. Up to now, using LLMs in historical research mostly meant summarizing PDFs or transcribing OCR text. Here, the system behaved like an autonomous junior investigator: when it lacked evidence, it cross-checked archives; when it lacked tooling, it wrote its own simulator; when inputs were noisy, it reasoned probabilistically.

For the broader tech space, this shifts how we think about solving long-tail historical problems. Millions of archival documents and cold cases sit untouched worldwide simply because the labor required to investigate them exceeds human bandwidth. If an AI agent can ingest messy records and build custom analytical engines on the fly, long-tail academic research just got a massive speed boost.

Of course, there is a catch. The human expert in the loop remains indispensable. AI can execute parallel investigations at blistering speed, but without experienced domain specialists setting boundaries and feeding the right context, it can easily hallucinate or chase rabbit holes. Still, turning months of tedious archival legwork into a 10-hour sprint is an incredible milestone.