a16z and Stanford Professor Discuss AI's Impact on Cryptography

Author: a16z crypto

Compiled by Deep潮 TechFlow

DeepInsight Summary: OpenAI’s latest mathematical model breakthrough has sparked industry concern, with some fearing that AI could dismantle the cryptographic systems we rely on. This article invites Stanford cryptography professors and the a16z crypto research team to clarify what is fact versus speculation, and to highlight the real opportunities AI may bring—a must-read for every professional dependent on cryptographic infrastructure.

Following OpenAI’s mathematical breakthroughs that prompted warnings from industry experts about the potential for AI to break the cryptography we rely on, we invited cryptographers from Stanford University and experts from a16z crypto to discuss what’s fact versus speculation, what truly warrants concern, and what doesn’t.

Dan Boneh, a professor at Stanford University, leads the Applied Cryptography group and co-directs the Computer Security Laboratory; he also serves as a special research advisor at a16z crypto.

Eddy Lazzarin, General Partner at a16z crypto

Justin Thaler, Research Partner at a16z crypto and Associate Professor of Computer Science at Georgetown University

Liz Harkavy, Investment Partner at a16z Crypto

Sonal Chokshi, Editor-in-Chief at a16z Crypto

Additionally, six key conclusions:

Recent results have demonstrated impressive advances in mathematics and computer science, but none of them include achievements in cryptography.

“This is a remarkable advancement in mathematics and computer science, but there’s nothing here about cryptography. … So you might wonder, does this mean these models are truly bad at cryptography? We don’t know.” — Dan Boneh, 03:00

We have not learned that elliptic curves are any more vulnerable than before. We have not learned that any specific cryptographic assumption is less secure than before. — Eddy Lazzarin, 15:19

A large part of modern cryptography is built on the assumption that certain mathematical problems have no efficient solutions. This may hold true just as well for increasingly capable AI.

We generally believe that these problems have no efficient algorithms. It’s not that we think they might exist but humans simply aren’t smart enough to find them—we believe they simply don’t exist. If they don’t exist, no matter how intelligent AI is, it won’t be able to find an algorithm that isn’t there. — Justin Thaler, 16:12

We must assume that certain problems are hard. We cannot prove these problems are hard because it relates to a famous unsolved problem in computer science: the P versus NP problem. In a sense, we must first solve this problem to truly argue that certain cryptographic problems are indeed, truly difficult. … Any problems released this week are nowhere near solving the P versus NP problem. We still need to keep making assumptions in cryptography, right?”, Dan Boneh, 06:00

Attempting to break cryptography is itself one way researchers build confidence. A design that withstands sustained, serious attacks earns trust. AI may change the speed at which this scrutiny occurs.

In the past, when someone proposed a hypothesis and built a system around it, we often had to wait about ten years to deploy the system, giving people enough time to analyze the hypothesis and build confidence. … Now, suddenly, we can use these models to dramatically accelerate this process. … We can now run the model for a month, trying to break it; if the model holds up, our confidence in the validity of the hypothesis increases significantly. — Dan Boneh, 12:10

Although recent concerns have centered on AI, some of these systems have already been targets of migration efforts due to the threat posed by quantum computing. This overlap has its limitations.

Due to concerns about quantum threats, elliptic curve cryptography is increasingly being viewed as something that may soon become vulnerable. Therefore, precisely because of these same concerns, we have already engaged with, prepared for, and deployed many of the mitigation measures you need—just originally intended for a different threat.” — Eddy Lazzarin, 20:52

“The threat from AI is, in some ways, broader, because it could also examine our post-quantum systems and weaken them in some way, or even fully break them,” Dan Boneh, 22:08

Cryptography agility refers to a system’s ability to replace its cryptographic methods when evidence or requirements change. It enables engineers to respond to unforeseen discoveries.

The system we deployed allows us to quickly switch from one design to another if the design in use is proven insecure. Cryptographic agility is, in a sense, our friend when it comes to defending against unforeseen circumstances,” Dan Boneh, 22:47

Dan pointed out that internet connections are already combining different cryptographic methods:

They are already using post-quantum cryptography. They are encrypting using both elliptic curves and lattices. Even if you break the elliptic curve, it’s still encrypted under a lattice-based system,” Dan Boneh, 38:40

One of the clearest opportunities for builders is to incorporate more security testing during development.

When discussing complex cryptographic systems, Justin pointed out:

The attack surface has become very large, and yes, I’m very optimistic that the AI tools causing people to panic about cryptography will also help us eliminate all these vulnerabilities. It’s a very difficult problem.” — Justin Thaler, 28:56

After the system is deployed, it will attract significant attention from cryptographers, but by then it will already be too late to attack the cryptographic system. Now, because we control these AI models, any startup deploying new cryptographic capabilities—or any company deploying any software—can use these models to analyze its cryptography and identify vulnerabilities throughout its software. — Dan Boneh, 35:05

Vulnerabilities can be eliminated. They can be avoided through design, mitigated, and the primary resource we use to mitigate them—AI capabilities—is dropping significantly in cost every few months.,” Eddy Lazzarin, 37:10

The quoted content has been slightly edited for clarity. Ellipses indicate omitted paragraphs.