Yesterday, November 20, 2026, Levent Alpöge, a mathematician at Anthropic, announced that he has found a proof of the Riemann Hypothesis with the help of Claude, Anthropic’s large language model. Proposed by Bernhard Riemann in 1859, the hypothesis is one of the seven Millennium Prize Problems that the Clay Mathematics Institute selected in 2000 as among the deepest open questions in mathematics, each carrying a $1 million prize. The announcement follows OpenAI’s claim in September 2026 that one of its models had resolved another of the seven, the Navier–Stokes existence and smoothness problem.
Neither claim has yet been definitely confirmed. Still, the direction is striking. Until now, the only Millennium Problem to have been solved was the Poincaré Conjecture, by Grigori Perelman in the early 2000s. If both new claims hold up, two of the three solved problems will have been cracked with the help of AI.
The speed at which AI has acquired mathematical ability is extraordinary. When ChatGPT was released in November 2022, it famously stumbled over elementary arithmetic. Less than three years later, in July 2025, AI systems reached gold-medal level at the International Mathematical Olympiad. By 2026, they are contributing to original research. We may have reached a mathematical inflection point: a period in which mathematical discovery becomes scalable in a way it never was before.
How did this happen? Language models were originally trained to predict the next word, not to reason. But mathematics has one advantage over ordinary language: whether an answer is right can be checked objectively. That makes it well suited to reinforcement learning, in which a model is rewarded for reaching correct results. A second step came with “reasoning models” from 2024 onwards, which spend more computation working through a problem, trying alternative approaches and correcting their mistakes before giving an answer. A third was coupling AI with formal verification tools such as Lean, which check every logical step of a proof mechanically. Both of this autumn’s headline claims came with such formal proofs.
This is wonderful news for science. It may be less welcome news for public-key cryptography, one of the foundational technologies of the digital economy. Invented in the 1970s, public-key cryptography solved the problem of establishing digital trust at scale. It lets parties verify identities, messages, software and transaction instructions without first having to exchange secret keys, and it underpins the internet economy, blockchain systems and the emerging economy of AI agents. An attacker able to break it could impersonate legitimate servers and software, forge digital signatures and authorize fraudulent transactions.
The security of public-key cryptography rests on a remarkable assumption: that certain mathematical problems are extremely hard to solve. Nobody has proved that they are. In other words, the security of much of the digital economy, from internet communication and electronic payments to digital identities and software updates, depends on the current limits of mathematical knowledge and computing power.
So far, the hardness assumptions behind public-key cryptography have withstood the efforts of the world’s best mathematicians and cryptanalysts. But they are already known to fail in a different model of computation. In 1994, Peter Shor showed that a sufficiently large quantum computer could efficiently solve the problems on which today’s public-key schemes rely. Advances in AI now raise the possibility that new mathematical methods could weaken these schemes even before powerful quantum computers arrive.
Quantum computers are not universally faster at mathematics. They excel at problems with a particular hidden structure that quantum algorithms can exploit. Post-quantum cryptography is built on different problems for which no such shortcut is known, on either classical or quantum computers. This may also make them more resistant to AI-assisted cryptanalysis, since an AI, too, must discover exploitable mathematical relationships to construct an efficient attack. But this is not guaranteed. AI could uncover mathematical structures or algorithms that nobody knows today. In other words, moving to post-quantum cryptography does not replace assumed security with proven security. It replaces one set of unproven assumptions with another that appears more resistant to the attacks we currently know.
A cryptographic scheme can also be broken without anyone solving the hard problem it rests on. In 2022, SIKE, a candidate that had reached the fourth round of the post-quantum standardization process run by the US National Institute of Standards and Technology (NIST), was broken in about an hour on a single ordinary computer. And in July 2026, Anthropic reported that its Claude Mythos Preview model had found a significantly improved attack on HAWK, a post-quantum signature scheme under evaluation by NIST, exposing a flaw in its mathematics that earlier expert review had not found. More generally, AI can help find flaws in implementations, protocols and key generation, as well as mathematical weaknesses in individual schemes.
These are uncomfortable times for public-key cryptography. Quantum computing is a known mathematical threat with an uncertain timeline, AI is an uncertain mathematical threat with rapidly growing capabilities. The prudent response is twofold. First, migrate to post-quantum cryptography now. Second, build infrastructure flexible enough to switch algorithms again if today’s choices prove vulnerable. This is what cryptographers call “crypto-agility”.
There is a further reason not to wait. Migration takes time, because it is far more than swapping one algorithm for another. Post-quantum algorithms typically use larger keys and signatures and place different demands on computation and bandwidth. The financial sector faces particular challenges because cryptography is woven through interconnected institutions, payment systems and settlement infrastructure. Changing the signatures used to authenticate interbank payment instructions, for example, may require coordinated upgrades by the payment-system operator, participating banks, technology vendors and service providers. The cryptography itself may be efficient, but ensuring that hundreds of institutions can adopt it without disrupting critical payment services is a much larger operational challenge.
Riemann’s hypothesis waited 167 years for a proof that may now have arrived in a matter of weeks. The mathematics protecting our payment systems may not get 167 years either. A mathematical inflection point is, by definition, the moment when the old pace stops being a guide to the future. For cryptography, that moment is now.