This paper argues that AI can accelerate brute-force cryptanalysis by learning patterns in the random-looking plaintexts generated by wrong decryption keys, thereby "flattening" the remaining key-space probability curve and speeding up key search. The author claims this threatens modern cryptography's assumption that trying wrong keys yields no information, suggests that NIST post-quantum crypto (PQC) is not immune, and proposes "Pattern Devoid Cryptography" as a defense—using non-trivial ciphertexts, unilateral randomness, and variable key sizes to deny AI pattern-recognition opportunities.
Main takeaways:
- Claims AI can find patterns in the "wrong plaintext" candidates produced by incorrect keys during brute-force search, making key search faster.
- Argues this violates cryptography's core assumption: "not learning from mistakes" (i.e., wrong keys should yield no information).
- Suggests NIST post-quantum cryptography (PQC) is vulnerable to this AI-accelerated brute force.
- Proposes "Pattern Devoid Cryptography" as a defense: ciphertexts should be non-trivial, use unilateral randomness, and support variable key sizes.
- Calls for a thorough review of cryptographic security posture in light of AI's pattern-recognition capabilities.