Can AI Break Cryptography? What Is Claimed vs. What Is Verified

Short answer: nobody has shown that an AI model can break any widely used cryptographic scheme. What exists is a claim, relayed by computer scientist Scott Aaronson on October 7, 2026, that AI companies have "started, gingerly and discreetly" testing whether their newest internal models can. Ethereum researchers took that seriously enough to call for precautions. This post sets out who said what, what is verified, and what is not.

It sits directly after OpenAI's release of hundreds of AI-written math manuscripts, which we covered in OpenAI's 722 math manuscripts: what to check and the eight results that matter. If you want the background on why labs may sit on results, see our earlier post on the rumor that labs are hoarding solved math problems.

TL;DR: claim, evidence, status

What did Aaronson actually write?

Aaronson's post "The Mathocalypse" (published October 7, 2026 on Shtetl-Optimized) is mostly a reaction to OpenAI's math release. He lists results he finds striking, including a claimed proof of the Unique Games Conjecture with a Lean certificate, and says the model was tried on roughly 8,000 problems and solved about 5% of them, with around three hours of compute per problem. He also contrasts OpenAI's habit of releasing raw proofs with the approach of having humans write a digested version.

The cryptography point is an update appended to the post. He observes that cryptography is "extremely conspicuous by its absence" from OpenAI's list of papers. He then writes that "my sources tell me that the AI companies have now started, gingerly and discreetly, investigating whether their latest internal models can break" important cryptographic protocols and primitives, and adds that if they can, it would be good to get ahead of it before the rest of the world works it out.

Three details matter for how much weight to give this:

• It is anonymous. No lab, person, model or date is named.
• It reports an investigation, not a result. The sentence is conditional throughout: "if they can".
• The headcount differs inside the post. The body speaks of 372 results while the update cites 376 papers, a reminder that even the numbers around this release are still being tallied.

One more caveat: the post's epigraph is a quote from Omer Reingold about complexity classes, shared with permission. It is not the origin of the cryptography claim, so be wary of summaries that attach his name to it.

Why would labs test this at all?

It is a natural thing to try. Cryptographic security is built on a short list of mathematical assumptions: factoring and discrete logarithms for RSA and elliptic curves, hardness of lattice problems for the newer post-quantum standards, and the behavior of hash functions. A model that can discover real new mathematics, as OpenAI says its does, is also a candidate tool for cryptanalysis.

There is also a responsible-disclosure logic. If a lab found a weakness, the usual norm in security is to notify maintainers before publishing. Aaronson's own phrasing hints at this: getting ahead of things before the rest of the world works it out. That is consistent with how labs now handle cyber capability. Anthropic, for example, tiers access to its cyber-capable models; see our coverage of the Anthropic cyber verification program and of AI attack swarms finding zero-days.

None of that proves anything has been found. It explains why the claim is plausible to experts without being evidence.

What did Justin Drake and Vitalik Buterin say?

According to Cointelegraph's report on Drake's call, Drake posted on October 7 asking the blockchain industry to "calmly begin planning for 'bunker mode'". His proposal is a controlled migration of assets, starting with large and sophisticated holders, to fresh addresses whose public keys have never appeared on chain. The reasoning: an address that has never signed a transaction exposes only a hash of its key, so an attack on the signature scheme (ECDSA) has nothing to work backward from.

Buterin replied on X the same day. As reported by The Block, he did not recommend rushing funds to new wallets, but said keeping assets in addresses that have never signed a transaction makes sense if it is easy, and that "botched migrations" have cost him more than hacks. He also said there is a good chance the concrete security of lattices "will take serious hits from the next two years of AI math", naming ML-DSA and fully homomorphic encryption as areas of concern, and said the industry should prefer hash-based constructions where possible. These are his opinions, not findings.

Note what neither researcher claimed: that a break exists. Both framed the issue as risk management on a timescale of months to a couple of years. Several secondary outlets, including aggregators, compress this into headlines like "AI may break cryptography within 2 years"; treat that wording as a paraphrase of a risk assessment, not a finding. This is a security and mathematics story, and we are deliberately not covering market reactions.

Claimed vs. verified

How would we know if the claim were true?

Cryptographic breaks tend to become public in recognizable ways. Watch for:

1. A preprint or lab post describing an attack, ideally with parameters and a reproducible script. The OpenAI math repo is the template: manuscripts plus machine-checkable artifacts.
2. A NIST or IETF statement. The post-quantum standards process is public. An advisory about ML-KEM or ML-DSA would be the strongest sign.
3. Independent cryptographers reproducing it. One lab's claim about its own model is not enough, as the Navier-Stokes dispute showed in our explainer.
4. Verification norms. The AGMAI responsible-release proposal is relevant: it argues for how AI-generated math should be released and checked.

Absence of these is exactly why this belongs in a "claimed vs. verified" post rather than a news flash.

What builders can do now

You do not need a migration plan on Thursday. You need visibility.

• Inventory public-key use. TLS certificates, SSH keys, code-signing, JWT signing, wallet and custody keys. Know which algorithm each uses.
• Aim for crypto agility. Systems that can swap algorithms by configuration are much cheaper to defend than ones with RSA-2048 hard-coded in a firmware image.
• Prefer conservative primitives where you choose. Hash-based signatures rest on fewer assumptions than lattices, which is the direction Buterin described, at the cost of larger signatures.
• Do not rush key rotations. Buterin's warning about botched migrations applies to ordinary teams too. Rehearse, stage, and keep rollbacks.
• Treat AI-generated attacks as a defender tool too. The same models can audit code and protocols; see our guide to an open-source security audit skill for agents.

How this connects to the bigger picture

Two trends are meeting. First, frontier labs are showing models that produce research-grade mathematics at scale. Second, labs are deciding what to publish and what to hold back for safety reasons. Cryptography is where those two collide most sharply, because a result that is a triumph in a journal is an incident in a production system.

It also changes how we should read silence. Aaronson's observation that cryptography is missing from the list is itself only an observation: it could mean nothing was found, nothing was tried, or something was found and withheld. The public record cannot distinguish these. That is the honest state of knowledge today, and it is why we label the lab-testing claim unverified.

Open questions

• Which labs, if any, are doing this testing, and under what disclosure policy?
• Would a lab publish a break, or notify standards bodies privately first?
• How does AI-found cryptanalysis compare to the quantum timeline that migration plans are built around?
• Will the 5% solve rate on open problems carry over to cryptanalysis, where problems are narrow and heavily studied?

We will update this post if a lab, NIST, or independent cryptographers say anything on the record.

Facts and quotes are accurate as of October 8, 2026, and come from the cited sources. The cryptography claim is anonymous and unconfirmed.

Related reading

• OpenAI's 722 math manuscripts: what to check
• OpenAI's eight headline math results and their Lean status
• Are AI labs hoarding solved math problems?
• AI did not solve Navier-Stokes: a PhD explainer
• AGMAI: responsible release of AI-generated mathematics
• Anthropic cyber verification program
• Armadin AI attack swarms and zero-days