In short, Google Brain researchers have discovered that the AI, when properly tasked, create oddly inhuman cryptographic schemes and that they’re better at encrypting than decrypting. The paper, “Learning to protect communications with adversarial neural cryptography,” is available here.
The rules of the task were simple. Two neural networks, Bob and Alice, shared a secret key. Another neural network, Eve, was tasked with reading the communications between the two robots. There was one condition, a “loss function,” for each party. Eve and the recipient Bob’s plaintext had to be as close to the original plaintext as possible while Alice’s loss function depending on how far from random Eve’s guesses were. This created a generative adversarial network among the robots.
“Informally, the objectives of the participants are as follows. Eve’s goal is simple: to reconstruct P accurately (in other words, to minimize the error between P and PEve). Alice and Bob want to communicate clearly (to minimize the error between P and PBob), but also to hide their communication from Eve. Note that, in line with modern cryptographic definitions (e.g., (Goldwasser & Micali, 1984)), we do not require that the ciphertext C “look random” to Eve. A ciphertext may even contain obvious metadata that identifies it as such. Therefore, it is not a goal for Eve to distinguish C from a random value drawn from some distribution. In this respect, Eve’s objectives contrast with common ones for the adversaries of GANs. On the other hand, one could try to reformulate Eve’s goal in terms of distinguishing the ciphertexts constructed from two different plaintexts.”
Cryptography Related Books
Note: I find it quite ironic (and to a degree funny) that the GCHQ publishes "entertainment" books, considering how serious they otherwise are about data and encryption.