GPT-6 Astra Cracked a 1941 Enigma Message and a 1918 WWI Cipher

Two separate builders reported GPT-6 Astra solving historical ciphers that had sat unsolved for decades this week — a 1941 German Army Enigma message and a 1918 WWI German naval radio transmission, both pulled from a public list of roughly 50 historically unsolved encrypted messages. One result got direct sign-off from a working Enigma historian; the other comes with an honest, still-unresolved question raised by the poster's own reply thread. Here's what actually happened in each case, and what to weigh before treating either as a settled result.

TL;DR

The Enigma result: a real historian signed off

Builder Carter Leffen posted that GPT-6 Astra broke a previously unsolved 1941 German Army Enigma message over roughly 10 hours, working autonomously through a genuinely multi-step process: searching historical archives, comparing uncertain letters in the source material, finding contextual clues, building its own Enigma simulator from scratch, writing cryptanalysis code, running parallel experiments testing competing keys, and cross-checking its recovered plaintext. The decoded message — labeled MVUEH, 82 characters — revealed a soldier stationed in Rosenow requesting marching route instructions, a mundane wartime logistics message rather than anything dramatic, which is itself a point in favor of the result's plausibility: a fabricated or hallucinated "solution" would more plausibly read as something more narratively satisfying than a routine request for directions.

What separates this from the usual unverified AI-capability claim explainx.ai treats skeptically by default — the same caution applied to GPT-6 Astra's unverified Portal and three.js claims — is that Leffen states he specifically sought and received review from Frode Weierud, a recognized Enigma historian and cryptologist who runs cryptocellar.org, before publishing. That's a real credentialed third party checking the work, not just the poster's own assertion. When one commenter directly challenged the claim as "a demo claim, not cryptanalysis evidence" without reproducible logs, Leffen responded by publishing a dedicated website with the full process log and a working three.js-based virtual Enigma machine — a substantive, verifiable answer to a fair challenge, not a dismissal of it.

The WWI result: verified against naval logs, with an honest open question

Separately, X user @deredleritt3r ("prinz") reported GPT-6 Astra deciphering a 1918 WWI German radio transmission — one of roughly 20 WWI-era German radio messages appearing on a public list (maintained at scienceblogs.de) of the top 50 unsolved historical ciphers. The decoded message translates to "an English cruiser arrived at Sevastopol on the [24]th, an allied squadron follows on the 26th." Astra reportedly cross-checked its own answer against real historical naval records and found that HMS Canterbury did report arriving at Sevastopol on November 24, 1918, with an allied squadron following on November 26 — an exact date match to the decoded content, and a genuine internal consistency check rather than a self-reported claim alone.

The most important detail here, though, comes from the poster's own follow-up reply, which deserves more weight than the headline claim: he notes that the encryption key used for this class of WWI message has been publicly available for a long time, and that many similar messages using the same key have already been solved in the past — which raises a real, still-open question about why this specific message sat unsolved if the tools to crack it already existed. A separate commenter flagged a genuine chronological oddity worth noting too: the message itself was sent November 27, but the key was allegedly only in use starting December 9 — a discrepancy nobody in the thread, including the original poster, fully resolved.

The skepticism worth taking seriously

Two specific objections in the replies deserve more than a dismissal as generic hater noise, because they're concrete and checkable. One commenter pointed out that multiple Enigma solvers already exist publicly online, raising a fair question about whether Astra effectively found and adapted an existing published solution rather than solving the cipher through cryptanalysis from first principles — a distinction that matters a great deal for how impressive the result actually is, and one the thread doesn't fully settle either way. A second commenter asked whether the results themselves are now on the internet in a way that could "contaminate" future model training or evaluation — a real, structural concern for any capability claim published publicly, similar to the eval-leakage problem explainx.ai covered in Exa's Snapshot launch earlier this week.

Worth noting on the other side: when directly challenged for reproducible evidence, Leffen didn't retreat to vague reassurance — he published logs and a working demo specifically to let others check the work themselves. That's a meaningfully different posture than most unverified viral AI-capability claims, and it's the reason this post treats the Enigma result with more confidence than the WWI one.

Why "already-known key" doesn't automatically mean "trivial"

It's worth sitting with the WWI result's unresolved question a bit longer, because it's genuinely ambiguous which direction it cuts. Having access to the correct daily key for a Kriegsmarine or Army cipher doesn't automatically make decoding a specific message trivial — WWI-era German military cryptography still required correctly identifying which cipher system and key schedule applied to a given message, handling transcription and OCR errors in a century-old handwritten or typed original, and getting the underlying substitution or transposition procedure exactly right before a candidate key even becomes testable. It's entirely possible this specific message went unsolved for a century simply because nobody with access to the right key ever got around to it, rather than because it was cryptographically harder than its already-solved siblings — an unglamorous, mundane explanation that's also the most likely one for a huge share of "unsolved" historical ciphers generally.

That ambiguity is itself worth naming plainly: neither "Astra did something impressive by rediscovering an available method" nor "this was a career footnote nobody prioritized" is confirmed by anything in the public thread. Readers evaluating how much credit belongs to the AI system specifically, versus how much credit belongs to the message simply being low-priority and previously unattempted, don't currently have enough information to settle it either way — which is exactly why this post treats the WWI result with meaningfully more hedging than the Enigma one.

Honest limitations

• Neither result is an OpenAI-published or peer-reviewed finding — both are independent builders' own reports, with varying degrees of third-party verification behind them.
• The Enigma result has real expert review (Frode Weierud) behind it; the WWI result does not, beyond the poster's own internal cross-check against historical naval logs.
• The "did it crib from existing online solvers" question for the Enigma result is not fully resolved by anything in the public thread.
• The WWI cipher's own poster flags an unexplained discrepancy — why a message using an already-public, already-used key sat unsolved for over a century — that neither he nor Astra has answered.
• No compute-cost comparison to brute-forcing the same problem was provided — a fair question raised directly in the replies and left unanswered.

What this means for builders

The genuinely interesting part of both results, independent of how much credit belongs to novel cryptanalysis versus pattern-matching against existing published material, is the workflow: autonomous archive search, building a purpose-specific simulator from scratch, writing and running cryptanalysis code, testing competing hypotheses in parallel, and cross-checking results against independent historical records — all without a human directing each individual step. That's the same shape of multi-hour, self-directed technical work explainx.ai has tracked in Claude's autonomous kernel optimization for biomolecular models and Qwen3.8-Omni-Flash's self-directed model-optimization experiment — a genuine, recurring pattern worth tracking regardless of the domain, more than any single historical puzzle solved.

Related on explainx.ai

• GPT-6 Astra reportedly beat Portal and wrote a Bach chorale — unverified
• GPT-6 Astra's launch: every benchmark, the pricing, and the ARC-AGI controversy
• Top 10 things to build with GPT-6 Astra
• Claude sped up 30 biology models 4x with FlashPairformer
• Exa Snapshot: search the web as it looked on any past date
• How to read AI benchmark claims critically
• Primary source: Carter Leffen's Enigma thread on X · MVUEH solution site
• Primary source: prinz's WWI cipher thread on X · The Top 50 Unsolved Encrypted Messages

This post is sourced to two independent builders' own X threads, published September 17-18, 2026, plus the linked supporting materials each provided. The Enigma result carries direct review from cryptologist Frode Weierud; the WWI result rests on the poster's own historical cross-check and carries an unresolved question the poster himself raised. Neither is an OpenAI-published or independently peer-reviewed finding.