Cryptography specialists report a deluge of amateur GPT-6 Astra cipher solutions, often wrong
Cryptography specialists say they are overwhelmed by requests to validate amateurs' GPT-6 Astra cipher solutions, Nature reports. A claimed solution can be ready by morning, and in Nature's two examples each went to one named person for confirmation.
The Scientist · Science desk
Key figures: Enigma's rotors and plugs could encrypt letters in up to 103 sextillion possible ways. The message GPT-6 Astra chose was 82 characters long. It had been encrypted 85 years before Leffen asked the model to decipher it.
| Figure | Value | Note | Claim |
|---|---|---|---|
| Ways Enigma could encrypt letters, up to | 103 sextillion ways | 3 | |
| Chosen message's length | 82 characters | 4 | |
| Encrypted years earlier | 85 years | 1 |
What happened
- In September, Carter Leffen, a product development coach at Bloomberg Law, had the model choose and decode an unsolved Enigma message from Frode Weierud's list of historical ciphers in Oslo.
- The model chose an 82-character message sent on 10 July 1941, possibly because a related message from the same day had been decoded in 2017.
- The message used Enigma settings different from the 2017 one, an unusual case that might explain why it stayed unsolved for so long.
- Days later, SentinelOne engineer Carter Church gave the model a picture of an 1809 cipher; it found a partial key online and reached a working key in about six hours.
Why it matters
- What it limits constraint A plausible-looking decode is not a verified one, so each claim needs a knowledgeable person to rule out a false positive before it counts.
- What's now possible capability A person with no cryptanalysis training can hand a model an image or a site listing and get a candidate solution within hours.
- What doesn't add up contradiction Nature's headline credits amateurs with cracking long-secret codes, but Church says the 1809 message could have been solved without AI and describes the gain as saved time and effort.
Start with the size of the search. Enigma allowed up to 103 sextillion settings [3], and GPT-6 Astra tested more than four billion of them in ten hours [5]. That is roughly four in every 100 trillion [13]. A slice that small means the model was narrowing the search with something, and the source describes it looking for settings that would reproduce a crib from the message solved in 2017 and also make the new message comprehensible [12].
Checking is the slow end. Nature reports that specialists welcome the surge of interest in codebreaking but are overwhelmed by requests to validate AI solutions, which can often be incorrect [14]. "There are so many false positives that are coming in from people who don't really understand cryptanalysis," said Elonka Dunin, a retired game developer who has written books on cryptography [15].
The checkers named in the piece are an independent cryptographer in Oslo, Frode Weierud [2], and Satoshi Tomokiyo, a patent engineer and hobby cryptographer in Tokyo [7]. Each received one amateur's solution for confirmation [16][9].
The thing this doesn't tell you is the denominator. Nature describes two successes. It does not give the number of validation requests, the share that are wrong, or how long a check takes.
In our view, the strain Dunin describes follows from a mismatch in effort. A candidate was ready for Leffen by morning [1], and judging it took a person. That is an inference from two cases and one quotation; a count of requests and a time per check would test it.
What to watch
- Any published confirmation from Weierud or Tomokiyo of the two solutions sent to them.
- A count from Dunin or the site operators of validation requests received and the share found incorrect.
- Whether reference sites such as Crypto Cellar Research or Cryptiana change how they accept solutions.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence40
- Adoption
- Not scored
- Hype gap+10
- Incentives
- Not scored
- Confidence55
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Carter Leffen, who has no cryptography experience, asked GPT-6 Astra to decipher a secret German army message encrypted 85 years earlier, nudged it a few times, and encouraged it to keep working when he went to bed; when he woke up in the morning, it had a solution.
- [2]
In September, Leffen, a product development coach at Bloomberg Law in New York City, asked GPT-6 Astra to scour Crypto Cellar Research, a website listing unsolved historical ciphers run by independent cryptographer Frode Weierud in Oslo, and to select and decode an Enigma message.
- [3]
Enigma's rotors and plugs rerouted electrical signals to encrypt the letters of the alphabet in up to 103 sextillion possible ways.
- [4]
GPT-6 Astra chose a message sent on 10 July 1941, possibly because it was short (82 characters) and potentially related to another Enigma message sent on the same date that had been decoded in 2017.
- [5]
GPT-6 Astra wrote code to simulate an Enigma machine and, in ten hours, tested more than four billion settings.
- [6]
The message had been encrypted with Enigma settings different from those used on the already-deciphered message sent the same day; this is unusual and might explain why the code went unsolved for so long.
- [7]
Carter Church, a staff engineer at SentinelOne, chose an unsolved cipher sent by an Italian official to one of Napoleon's generals in 1809, listed on Cryptiana, a website compiled by Satoshi Tomokiyo, a Japanese patent engineer and hobby cryptographer based in Tokyo.
- [8]
Days after Leffen's success, Church gave GPT-6 Astra a picture of the 1809 message; the model found a partial key among records on a French cryptography website, used simulated annealing, and hit on a successful key in about six hours.
- [9]
Church sent the decoded 1809 message, which turned out to describe troop locations, to Tomokiyo for verification.
- [10]
Church notes that the 1809 message could have been solved without AI, but the model's ability to transcribe the symbols in the image, find the partial key and determine how to apply it greatly reduced the time and effort required.
- [11]
Non-experts are using large language models to break historical codes, harnessing the models' ability to rapidly deploy decryption algorithms while searching the Internet for extra historical documents or information.
- [12]
GPT-6 Astra searched for the Enigma settings that would both reproduce the crib from the message solved in 2017 and render the unsolved message comprehensible.
- [13]
Four billion settings tested out of up to 103 sextillion possible is roughly four in every 100 trillion.
- [14]
Cryptography specialists are thrilled at the surge in interest in codebreaking but are being overwhelmed by a deluge of requests to validate AI solutions, which can often be incorrect.
- [15]
Elonka Dunin, a retired game developer in the Washington DC area who has written books on cryptography, said: "There are so many false positives that are coming in from people who don't really understand cryptanalysis."
- [16]
Leffen sent the model's solution to Weierud for confirmation.
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
1 independent publisher whose own reporting we read for this story.
- nature.comHistory’s long-secret codes are being cracked by amateurs using AI
2 articles · October 15, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.
Topics
- Large language modelsFollow
- CryptanalysisFollow
- AI codebreakingFollow
Entities
- OpenAIFollow
- Enigma machineFollow
- Satoshi TomokiyoFollow
- Frode WeierudFollow
- SentinelOneFollow
- Elonka DuninFollow
- GPT-6 AstraFollow