A week after OpenAI dropped 722 mathematical manuscripts written by an unreleased AI model, the cryptography community is arguing about a question most people had filed under science fiction: could an AI system eventually break the encryption that protects the internet? According to a report by Scientific American, the experts are genuinely split.
The debate stems from OpenAI's record-setting release, in which every paper was penned by a model the public cannot access. The output was far from flawless — OpenAI withdrew three of the papers within a day over a sign error, and mathematicians are still reviewing the rest — but the scale and incisiveness of the work rattled researchers who had assumed machine-generated mathematics was years away. We covered the initial release of the 722 manuscripts and the subsequent withdrawals as they happened.
The mathematics community's reaction to the release itself has been combative — leading mathematicians, including Fields medallists, have publicly demanded stronger verification norms for AI-authored proofs, and the cryptography debate is in many ways an extension of that argument into applied security. If AI-generated mathematics cannot yet be fully trusted, the question becomes which parts of the trust infrastructure can afford to wait.
The 15 Percent Worry
Mathematics is the foundation of cryptography. Every encrypted message and bank login relies on a math problem believed to be practically impossible to reverse. Matthew Green, a cryptographer at Johns Hopkins University, put a number on the worry in a social media post days after the release: he sees roughly a 15 percent chance the world could lose confidence in existing public-key encryption algorithms — the systems that let browsers establish secure connections and verify digital identities without sharing a secret key in advance.
The fear has historical precedent. Cryptographers once thought 512-bit RSA keys were safe, until the number field sieve made cracking them dramatically easier and forced the industry to lengthen keys — 2,048 bits is now the accepted minimum. Green's concern is that AI could find a similar shortcut in elliptic curve cryptography, which secures much of the web and most cryptocurrencies, by spotting structure humans have missed for decades.
The Skeptics: No Evidence of a Breakthrough
Plenty of leading cryptographers see no cause for alarm. Yehuda Lindell, a professor at Bar-Ilan University and head of cryptography at Coinbase, wrote that there is no evidence whatsoever of a breakthrough against the mathematical problems underlying modern encryption. Dan Boneh, a Stanford cryptography professor, pointed out that not a single one of OpenAI's 722 papers deals with cryptography at all.
Govind Pimpale, a research fellow at the Foundation for American Innovation, conceded the current state of play plainly: no currently deployed algorithm has been broken by AI. His concern is structural rather than immediate — public-key cryptography rests on a handful of math problems studied by small communities, and machines could throw vastly more effort at those narrow targets than human researchers ever have.
What the Evidence Actually Shows
The most concrete data point comes from a recent preprint by Florian Tramèr, a computer scientist at ETH Zurich, and colleagues including three researchers at Anthropic. The team pitted leading AI models against 191 code-breaking challenges, and Tramèr said it is clear that models' cryptanalysis ability is growing extremely rapidly.
Other researchers redirect the worry toward nearer-term risks. Eyal Ronen of Tel Aviv University, a co-author of the cryptanalysis study, noted that AI is already very good at finding software vulnerabilities — a more immediate threat than hypothetical encryption breaks, since attackers overwhelmingly take the easiest route available, such as stealing keys from a compromised machine. Bart Preneel, a cryptography professor at KU Leuven, added that scale matters: even a mathematical advance that made attacks a billion times faster could still leave them computationally impractical against properly configured systems.
Why the Stakes Are So High
Public-key cryptography quietly secures browser connections, software update signatures, cryptocurrency transactions and messaging apps. A genuine break would let attackers impersonate banks, forge software updates and read years of stockpiled encrypted traffic — the harvest-now, decrypt-later scenario intelligence agencies are assumed to already attempt against post-quantum transitions.
Even a partial result would force a painful global migration. Green warned that the upgrade process itself could be a catastrophe, noting that botched cryptographic transitions have historically introduced new weaknesses. That risk was on display just this month when Anthropic's Claude Mythos system identified previously unknown weaknesses in encryption implementations, a reminder that the same AI progress impressing mathematicians is also being aimed at security research.
For now, the consensus sits where it usually does in cryptography: no deployed system is broken, but the cost of assuming that lasts forever is too high to ignore. What changed this month is that the pace of mathematical discovery itself became the variable under surveillance.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
