Paper 2025/2161 Original: eprint.iacr.org/2025/2161 ↗

Attacks and Remedies for Randomness in AI (and with AI): Cryptanalysis of PHILOX and THREEFRY

Abstract

In this work, we address the critical yet understudied question of the security of the most widely deployed pseudorandom number generators (PRNGs) in AI applications. We show that these generators are vulnerable to practical and low-cost attacks by introducing both manual and automated cryptanalysis techniques. For this we partially, and to the best of our knowledge for the first time, take advantage of AlphaEvolve, a novel evolutionary coding agent guided by Large Language Models, to automate the discovery of complex cryptographic distinguishers. Finally, we present a cryptographically secure and well-understood alternative, which has a negligible effect on the overall AI/ML workloads. More generally, we recommend the use of cryptographically strong PRNGs in all contexts where randomness is required, as past experience has repeatedly shown that security requirements may arise unexpectedly even in applications that appear uncritical at first.