Ethereum creator Buterin calls for ‘post-AI cryptography’ amid rapid AI progress
Artificial intelligence’s rapid progress in mathematics has opened a split among cryptographers over whether the mathematical locks protecting the internet — bank accounts, email inboxes and everything else encrypted — could fail sooner than expected, potentially before quantum computers deliver the long-feared blow, according to a Wall Street Journal technology column by Christopher Mims.
The security of nearly all online communication rests on an assumption: that certain math problems are easy to run in one direction and essentially impossible to run in reverse. Multiplying two large prime numbers is simple; factoring the lengthy result back into its original primes is, with conventional computers, essentially impossible. For encryption to keep working, the column noted, the assumptions behind its algorithms must remain true — and some cryptographers now argue it may be unsafe to count on that.
Quantum computers have threatened that assumption for decades. Engineers have spent years preparing replacements, an approach known as postquantum cryptography, and according to leading mathematics experts cited in the column, an orderly global transition to those new algorithms should avert a failure that could otherwise set technology back a century.
The newer worry is that AI could get there first. The column reported that artificial intelligence is pushing advances across mathematics at an unprecedented rate, conquering problems humans have wrestled with for centuries, and that some cryptographers fear those advances could eventually reach the math problems that keep encrypted communications encrypted.
“A lot of what makes us feel like cryptography is secure is because a lot of smart people have looked at crypto for a long time, and we didn’t find anything,” said Matthew Green, a cryptographer and associate professor at Johns Hopkins University. “But what we’re learning is that human beings looking at things is not the gold standard—it’s not relevant anymore.”
Green pointed to recent mathematical breakthroughs by AI labs, including a partial solution to the infamously challenging Navier-Stokes equations and a more recent breakthrough disproving the Exponential Time Hypothesis. The latest of those breakthrough math results came from OpenAI and Anthropic, according to the column.
The Exponential Time Hypothesis is a long-held proposition that certain problems admit no shortcuts — the only way to solve them is to take the long way. If the hypothesis is false, the column noted, decades of research in theoretical computer science could be rendered irrelevant, and shortcuts to those problems could make cracking “uncrackable” encryption far easier. It is that result, the report said, that has some researchers especially worried.
Vitalik Buterin, creator of the Ethereum protocol that underpins much of modern cryptocurrency, voiced the concern publicly before those results appeared. “We should take the risks to cryptography from AI-accelerated math seriously, and minimize our exposure to not just quantum-vulnerable cryptography, but also potentially AI-vulnerable cryptography,” he wrote on X in early October. On Monday, he repeated the warning at the Ethereum Foundation’s annual conference in Mumbai, calling for “post-AI cryptography.”
The column likened the moment to the cryptographic race of World War II, when Alan Turing and his team at Bletchley Park — innovating on an electromechanical system pioneered by Polish engineers — found a way to crack Germany’s Enigma code, which many great minds had judged unbeatable.
Green cautioned that there is no evidence the world’s cryptographic algorithms have already been cracked. But he said he thinks the next AI-powered math breakthroughs could touch more directly on cryptography’s foundational math.
Yehuda Lindell, head of cryptography at Coinbase and a computer-science professor on leave from Israel’s Bar-Ilan University, argued that baseless speculation about an impending collapse of the world’s online security is not just premature but can be irresponsible. “Anything is possible,” Lindell said. “But speculating that bad things can happen is something which just doesn’t help anyone.”
Lindell objected in particular to claims that postquantum algorithms, which are gradually being implemented, might prove vulnerable enough to AI’s mathematical discoveries to force a rethink of the safeguards. He said the evidence points the opposite way: the new algorithms are likely to withstand most attacks by any kind of computer, and AI may end up strengthening cryptographic defenses. At worst, he said, engineers might have to lengthen cryptographic keys — the numbers used in secure exchanges — which would slow the constant encoding and decoding of online interactions without making them unusable.
The disagreement traces to fundamental mathematical questions that remain unresolved. Those who share Green’s concern worry that somewhere in mathematics sits an as-yet-undiscovered result that could make even some “impossible” problems solvable with today’s computers, weakening digital locks. Those on Lindell’s side believe decades of work by mathematicians suggest that possibility is vanishingly small. “My claim isn’t that AI will not help with cryptographic analysis—I think it very likely will,” Lindell said. He said he believes the assumption on which his field rests — that there really are hard problems that will never be solved by a shortcut — will stand the test of time.
Even in the worst-case scenario, Lindell said, he would rather an AI lab such as OpenAI or Anthropic be the one to discover modern cryptography is insecure, since those companies have a history of disclosing vulnerabilities so engineers can patch them. “We don’t want a system to be broken by North Korea, and then they don’t tell anybody else,” he said.
Green countered that because cryptographic systems take years to update, discreet disclosures would not be enough to protect them. “I definitely think that we don’t have any idea what’s coming,” he said.