Quantum Was the Deadline. Now Regulators Are Asking Whether AI Gets to Encryption First

For a decade, the threat to encryption had a name and a rough date: a large quantum computer, probably in the 2030s. Organizations were told to plan their move to post-quantum cryptography on that timeline. A report in the Financial Times on October 28, 2026, adds a second, less predictable threat. Mathematicians and regulators are asking whether fast-improving AI could weaken the mathematics behind today’s encryption before quantum machines arrive, and possibly the mathematics behind the post-quantum replacements too.

No AI system has broken a core cryptographic method. Several security experts quoted in coverage of the report are skeptical that one will soon. But the UK’s Financial Conduct Authority is not waiting to find out. It is trialling a supervisory tool with 50 large institutions to test their readiness for cryptographic risk, and plans to issue guidance next month to the more than 35,000 firms it supervises. Its message, according to coverage of the FT report, is that firms should start preparing now.

What set off the concern

The trigger was OpenAI’s release of hundreds of mathematical papers, including proofs of long-standing problems that it said involved AI-generated solutions. The release divided mathematicians; one group of them criticized it publicly. Whatever the merits of individual proofs, the episode made a point that cryptographers had been discussing quietly: AI is getting better at exactly the kinds of reasoning, in number theory and algorithm design, that underpin cryptography.

Much modern encryption relies on problems that are easy in one direction and hard in the other. Multiplying two large primes is quick; working out the primes from the product is, as far as anyone knows, extremely slow. Security rests on “as far as anyone knows.” Australian mathematician Geordie Williamson described AI’s effect to the FT as a “complete kind of restructuring of what difficulty is,” and said he expects major shocks. Daniel Shiu, of the University of Manchester, called AI a “significant risk” for “more mathematical forms of cryptography,” including “less mature” post-quantum algorithms.

Why post-quantum algorithms are part of the worry

This is the uncomfortable part. Post-quantum cryptography replaces today’s algorithms with ones built on different mathematical problems, such as lattices, that quantum computers are not expected to solve efficiently. Those problems have been studied for less time than factoring. If AI speeds up mathematical discovery, newer schemes may be exposed to new attacks faster than older ones were.

That does not mean organizations should slow their post-quantum migration. The quantum threat is still real, and “harvest now, decrypt later” attacks mean data encrypted today can be stored and broken later. It means the migration should be built so that algorithms can be swapped again if needed. Browser and platform vendors are already moving in that direction, as our coverage of Chrome’s quantum-safe certificate plan and of Google’s 2029 migration deadline showed.

What the FCA expects

The FCA’s published work gives a sense of where its guidance will go. Its Cyber Coordination Group insights for 2025 recommended that firms embed AI adoption and post-quantum migration in existing risk frameworks, supported by strong “cryptographic hygiene” and risk-based prioritization. It found that firms with good hygiene had cryptographic estates that were easier to map, clear ownership of key lifecycle processes, and vulnerable assets that were simpler to identify. That turned migration “away from a broad discovery exercise to a targeted, risk-led programme.”

The same report flagged third-party AI oversight as a concern: firms struggle to see how suppliers deploy AI, and without contractual visibility “risk inheriting unknown vulnerabilities.” It also noted a rise in attackers using AI for phishing, vishing and deepfakes.

Why this is a privacy issue, not just a security one

Encryption is one of the main safeguards privacy law relies on. Under the GDPR, encryption is named as an appropriate security measure, and a breach of properly encrypted data may not require notifying individuals if the key was not compromised. Many US state breach laws contain similar encryption safe harbors. If the strength of encryption becomes uncertain, so does the legal value of those safe harbors.

There is a second link. Long-lived personal data, such as health records, genetic data, financial histories and identity documents, is exactly what “harvest now, decrypt later” attackers want. Privacy teams that decide retention periods are, in effect, deciding how long data has to survive future cryptanalysis. Advocacy groups have pushed for wider encryption for years, as in the campaign we covered in EFF’s “Encrypt It Already” push. The new concern is about how long that encryption holds.

Six steps to take now

1. Inventory your cryptography. List where encryption is used, which algorithms and key lengths, who owns each key, and which systems protect personal data.
2. Rank data by how long it must stay secret. Data that must stay confidential for ten years or more should move first.
3. Build for crypto-agility. Design systems so algorithms can be changed without rebuilding the application. Assume you may have to change them twice.
4. Cut what you keep. Data you delete cannot be decrypted later. Retention limits are a cryptographic control.
5. Ask your suppliers. Add questions about post-quantum plans and algorithm agility to vendor assessments, especially for cloud, payments and identity providers.
6. Protect more than data at rest. Model weights, keys and backups are also targets. We discussed this in our piece on Illinois treating model weights as a security asset.

Australia’s cyber agency has published a useful primer on these risks, which we summarized in the ACSC quantum technology primer. Captain Compliance helps privacy teams keep records of the systems and data they hold, which is where a cryptographic inventory starts. We do not assess the strength of algorithms or certify encryption.

Frequently asked questions

Has AI broken encryption?

No. No core cryptographic method has been broken by AI. The concern is about the pace of AI progress in mathematics and what it might make possible.

Should we pause our post-quantum migration?

No. The quantum threat remains, and stored data is already at risk from future decryption. Build the migration so you can change algorithms again if needed.

When will the FCA guidance be published?

Reports say next month, aimed at the more than 35,000 firms the FCA supervises. Check the FCA’s website for the final document.

Do encryption safe harbors in breach laws still apply?

Yes, under current law. But they depend on encryption being effective. If an algorithm is weakened, regulators may question whether data protected by it was truly secured.

Cryptography has always been a bet that some problems are hard. The FT report is a reminder that “hard” is a statement about what we know, and what we know is changing quickly. Regulators are asking firms a simple question: if the bet stops paying, how fast can you change it? Most firms do not yet have an answer.