Another Historic Cipher Falls to AI
This one is from 1809, written by Napoleon’s nephew.
This one is from 1809, written by Napoleon’s nephew.
Jammy • October 5, 2026 10:00 AM
What happens now?
Clive Robonson • October 5, 2026 10:27 AM
@ Bruce, ALL,
We talk about security by obscurity
But not with encryption systems.
Now AI is getting to the point older codes and ciphers are getting broken we should think that “obscurity” is not the right word to use for codes and ciphers.
People should ask a question about both codes and ciphers especially of the older styles.
“What can not change through the whole process?”
And the answer is the “statistics” of the plaintext and ciphertext, otherwise you would not be able to decipher the ciphertext back to plaintext.
What we know is that Current AI LLM systems are built around so called “Digital Neural Networks”(DNNs) that are in reality just glorified “Digital Signal Processing”(DSP) systems configured as “adaptive filters”.
The thing about filters is they are very good at “lifting signals out of the noise”. Adaptive filters are very good at lifting signals out of background noise that changes constantly but in a predictable way.
Keytext in older ciphers often had a period the same length as the password/phrase used to generate the Keytext. And due to the failings of humans this was often quite short and mostly either around ten characters or less in length or a well known phrase. This remained true up untill and through WWII and was still used via hand ciphers by spys or simple mechanical machine ciphers up untill the 1980’s when the first microcontrollers with sufficient grunt became available.
Thus as plaintext has recognisable statistics even when combined with a keytext of recognisable ststistics (see chi-squared test).
We can expect Current AI LLM’s in an agenetic framework to basically pull out the desired signals.
Thus the only cipher systems we should expect to have some security are those where there is either no periodicity in the key text or the workload is many orders of magnitude greater than even large teams of humans can perform.
In short Current AI LLMs are going to be “on desktops” in the not to distant future, and with them your own “Pocket SigInt Agency”.
Subscribe to comments on this entry
Sidebar photo of Bruce Schneier by Joe MacInnis.
KC • October 5, 2026 9:40 AM
The announcement thread on x.
https://x.com/CarterWChurch/status/2105167567706816690
It is interesting that this recovery was just a time problem. See the write-up’s section 10: “A patient specialist could have done this in 1970 … “