GPT-6 Astra Solves a WWI German Radio Cipher
The post's own account makes this a known-key decoding with a documented key used earlier than expected, not a break of an unknown cipher.
The post's own account makes this a known-key decoding with a documented key used earlier than expected, not a break of an unknown cipher; the discussion's sharpest objections are about verification, misleading framing, and whether the LLM contribution is notable.
The post claims GPT-6 Astra solved a previously unsolved WWI German ADFGVX radio message from Nov. 27, 1918, yielding a German plaintext about an English cruiser arriving at Sevastopol and an allied squadron following on the 26th. It says the solution used the key TRUPPENVERSCHIEBUNG from J.
Rives Childs's work, explains the ADFGVX table and columnar transposition, and notes the key was documented as starting Dec. 9, 1918, two weeks after the message. The author says Astra checked the plaintext against HMS Canterbury's logs and found the cruiser arrived Nov. 24 and an allied squadron followed Nov. 26, and says he is unaware of prior decoding.
Key takeaways
- The article presents GPT-6 Astra's claimed solution to a Nov. 27, 1918 German ADFGVX radio message listed among unsolved ciphers.
- The claimed plaintext says an English cruiser arrived at Sevastopol on the ?4th and an allied squadron follows on the 26th.
- The claimed key is TRUPPENVERSCHIEBUNG, described in J. Rives Childs's work; the article walks through alphabetical key rearrangement and columnar transposition over 170 characters.
- Astra's stated hypothesis for prior failure is that the key was documented as starting Dec. 9, 1918, while the message was sent Nov. 27, 1918; the discrepancy is unexplained.
- The author says Astra checked the result against HMS Canterbury's original logs, which show arrival at Sevastopol on Nov. 24, 1918, and an allied squadron on Nov. 26, and says he is not aware of this message having been decoded before.
What people think
Verification and possible fabrication
iltk asked whether the ship logs are online, arguing that if so the model could have manufactured a fake key and matching message; stalfie pointed to the article's original log images, and JoshTriplett argued that a fabricated key is unlikely because the key was documented as used for other messages.
Known-key decoding, not brute force
Legend2440 said the key list was seized after the war, the message was missed because the operator mistyped the key and used a key from the wrong day, and ChatGPT only had to pick the correct key and identify the typo; msy asked whether this is just brute-forcing a small key; meindnoch said the only gotcha is the key being used two weeks earlier than documented; grey-area called the headline misleading.
Source verification challenged
eis said Astra could not describe pages 214-215 of the cited Childs book and could not find the book online; sehw suggested the model may have found a human solution in its training data and presented it as its own.
Technical correction on ADFGVX
FabHK asked whether ADFGVX is just a simple substitution cipher vulnerable to frequency analysis; jgrahamc replied that it also includes a keyed columnar transposition after the ADFGVX Polybius-square step. adverbly asked why TRUPPENVERSCHIEBUNG, an actual German word, could not be found by dictionary attack and sentence checking.
Goalpost-moving debate
j-pb said AI being as capable as a sufficiently dedicated human analyst is no longer considered noteworthy; roenxi argued any AI task is presumably within reach of a dedicated human or team; durdn satirized successive dismissals of LLM cryptanalysis results.
Broader LLM capability and context limits
aaymeloglu said agents are finding low-hanging fruit in unsolved ciphers, greenavocado noted the jaded reaction to calling 100-year-old unsolved ciphers low-hanging fruit, and znpy/irthomasthomas discussed whether larger useful context windows are the next frontier.
Estimates describe sampled comments, not the whole community.