Cryptography in the age of Mildly Superhuman AI
Zygi Straznickas
Abstract:
Two years ago, humans authored all cryptography research. It’s likely that soon, mathematical ASI will settle whether computationally secure cryptography exists. We're living in the transition – current frontier models are already expert-human-level, maybe superhuman, in mathematics and engineering, and occasionally break some cryptosystems ad hoc.
In this talk, we’ll discuss and speculate how cryptography could look during this period of mildly superhuman AI. We’ll cover both the present – how AI can already be used to aid cryptographic research in constructing primitives, drafting and checking reductions and verifying implementations; and the future – which cryptographic directions become newly relevant as intelligence becomes scalable on-demand, from end-to-end formally verified security reductions and implementations to understanding what confidence in hardness assumptions means when cryptanalysis is cheap and continuous.
Bio:
Zygimantas (Zygi) Straznickas is an AI safety researcher and systems engineer at Anthropic specializing in machine learning systems, white-box model evaluation, and adversarial robustness. His background spans theoretical computer science, formal mathematics, and post-quantum cryptographic lattice schemes. He did his Masters at MIT in 2020.