Guest Post by Ofir Joseph, Quantum Communication Researcher, University of Haifa, and Head of the Cybersecurity Program, Academic College of Tel Aviv–Yaffo, Prof. Irit Hadar, Dean, Faculty of Computer and Information Science & Director, Center for Cyber, Law and Policy, University of Haifa; Sean Galantzan, Researcher, Tel Aviv University and Miriam Shtilman, Tal Ventures
A bank does not need to lose a password for its security to weaken. A new mathematical method could make a previously impractical attack feasible. Research involving artificial intelligence (AI) is bringing that possibility into sharper focus.
Online payments rely partly on public-key cryptography, also called asymmetric cryptography: methods using two related digital keys, one public and one secret. These keys are strings of data, not passwords. They help computers establish private connections and check who issued information. Their security depends on certain calculations remaining too difficult for attackers.
Much preparation has centred on quantum computers, which exploit the unusual behaviour of matter at very small scales to perform some calculations differently. A sufficiently capable machine could defeat widely used public-key methods. But better mathematical methods could also change the odds, with AI helping researchers find them.
What OpenAI reported
On 26 October 2026, OpenAI reported that its AI had refuted the Exponential
Time Hypothesis, a longstanding, unproved conjecture about how computation grows as problems become larger.
Imagine yes-or-no decisions linked by rules. Each rule involves three decisions and is satisfied when at least one has the required yes or no answer. All the rules must hold together. Every additional decision doubles the possible combinations. The hypothesis does not require checking them all. It says that no general method can avoid exponential growth in the work needed for the hardest cases: roughly, repeated multiplication rather than a steady increase.
The released work describes a method that escapes that growth barrier and even counts the valid combinations. OpenAI says the result was checked in Lean, software that checks proofs expressed in a precise computer-readable language, while independent researchers examine it. This does not make every hard problem easy: a theoretical improvement can remain impractically slow.
A warning, not a collapse
Post-quantum cryptography, or PQC, uses mathematical methods designed to resist both ordinary and quantum computers. It runs on ordinary hardware. ML-KEM, one standardised method, establishes a shared secret used by a separate method to encrypt messages. Its security rests on a different mathematical problem. Turning a theoretical advance into a practical attack requires further work; the announcement does not demonstrate one.
Research published in July 2026 by AI developer Anthropic provides a more direct example. With AI assistance, researchers improved an attack on HAWK, a proposed post-quantum digital-signature scheme. A digital signature helps establish who issued information and whether it was altered. Researchers recovered a secret signing key in a small experimental version; attacking larger versions remained impractical.
Anthropic put the research effort at roughly 60 hours and $100,000 in model usage, not a price tag for breaking any chosen key. HAWK’s developers withdrew it on 29 July. NIST, the US standards agency, says the finding does not affect its finalised post-quantum standards. This is also defensive research working before a proposed method becomes a standard.
Separately, on 1 October, Germany’s cybersecurity authority, BSI, advised against choosing Classic McEliece, another key-establishment method, for new developments. It reported no practical attack against its recommended configurations. That was a different development, not evidence of another AI-driven break.
Mathematical attacks are not new: researchers have overturned promising cryptographic designs before, without AI. The emerging question is whether AI accelerates that process, for attackers and defenders alike.
Where quantum communication fits
Quantum key distribution, or QKD, offers another way to establish a shared secret. Two sites exchange specially prepared light signals and test some of their results. The results help limit what an eavesdropper could know about the resulting key. If the security conditions are not met, the procedure stops.
Its security proofs use quantum physics and explicit assumptions about equipment and operation. The key-distribution step need not depend on an attacker’s inability to solve a particular mathematical problem. A new shortcut against a public-key method does not, by itself, invalidate that basis. This makes QKD worth evaluating as a complement to PQC on selected sensitive links. A combined design must show what remains protected if one component fails.
The limitations are substantial. QKD needs dedicated equipment and suitable optical connections. It needs a separate way to verify the participants, possibly a secret installed beforehand. It neither replaces digital signatures nor protects a compromised computer. And supplying a key does not remove the need for secure encryption of the messages themselves. QKD is therefore not a direct remedy for HAWK.
The US National Security Agency does not currently recommend QKD for national security systems unless technical limitations are overcome. The UK’s national cybersecurity authority also rejects it for government and military use. The European telecommunications standards body ETSI, meanwhile, presents QKD as complementary to PQC. The case for combining them must rest on tested benefits, costs and operational risks, not a claim of immunity to AI.
What businesses should do now
Businesses should continue adopting standardised PQC and build the ability to change protection without replacing entire systems. The World Economic Forum’s work on quantum-safe migration makes a similar case for knowing where cryptography is used and preparing for change.
Readiness should also produce evidence: a map of which services depend on which protections, written supplier commitments, and tests showing that upgrades preserve security and essential operations. For a bank, payments must keep working while protections change. Buying quantum equipment is not, by itself, proof of security.
For sensitive links between data centres, independent trials could establish whether adding QKD provides worthwhile protection and what happens when a link fails. The aim is not to replace one promise with another, but to preserve options when the science changes.
The next challenge to encryption may come from a better idea, not just a more powerful machine. Businesses need to prepare for both.
