Is AI breaking cryptography?

Summary: On October 6, OpenAI published 722 mathematics manuscripts written by an internal model it hasn't released. Many of papers claim progress on problems that humans haven't been able to solve for decades.

It's important to note that one of the current mathematics work had anything to do with cryptography.

Nevertheless, the next day, Ethereum Foundation researcher Justin Drake told large crypto holders to start planning a calm move into addresses that have never revealed their public key, in case the math that protects their wallets falls sooner than anyone planned.

At the same time, Europol published a report naming exposed wallets as the weak point in crypto's security.

Thesis: Crypto's security rests on a math problem that has never broken. This week a machine showed how quickly "never broken" can expire.

Your wallet is guarded by a problem no one has proven hard

A crypto wallet is a pair of numbers that work together to keep the wallet accessible and useable, but also secure and in control of the user.

The private key is the secret. The public key is computed from it, and the math runs easily in one direction and, as far as anyone knows, can't be run backward, which is what keeps wallet balances secure.

Working out a private key from a public one would take today's computers longer than the universe has existed. This is a core value proposition of crypto and why it's considered so novel as a form of open money.

More specifically, Elliptical Curve Digital Signature Algorithm, ECDSA (the signature method Bitcoin and Ethereum use to prove you own your underlying assets), and it has protected essentially every crypto asset since 2009.

Cryptographers assume that ECDSA is bombproof because the best mathematicians alive have tried to reverse it for decades and failed.

There's a second layer to this story: Your address, the string you give someone so they can pay you, is a scrambled fingerprint of your public key rather than the key itself.

Reversing the scramble is a separate problem, and a much harder one by current estimates. So as long as you've only ever received money at an address, an attacker has two locks to pick.

The moment you send money out of that address, your signature reveals the public key, revealing aspects of the second lock.

On Ethereum that happens to every account that has ever sent a transaction. On Bitcoin it depends on the address type and on whether you reuse addresses.

A machine knocked over hundreds of problems in a week

OpenAI's repository says the company posed roughly 4,000 open problems to an unreleased internal model, and that "the vast majority" of what came back was produced by the same procedure.

It published 722 manuscripts grouped into 372 result families. The average result took about three hours of the kind of thinking compute a ChatGPT Pro subscriber can buy.

Martin Bridson, president of the Clay Mathematics Institute, said it would recently "have been impossible to imagine that the frontiers of mathematics could move so far."

The quality of the freshly published papers is uneven, "Some of the unformalized results could have issues," the README OpenAI warns. About 300 of the results carried proofs written in Lean (software that checks every step of a proof mechanically, the way a compiler checks code), but not all of them.

OpenAI withdrew three manuscripts over a single sign error right after publication. A company spokesperson said its own mathematicians don't yet understand many of the results.

What's interesting is that none of the papers so far address cryptography. Dan Boneh, who runs Stanford's applied cryptography group, said on an a16z crypto discussion published October 10: "This is incredible progress in mathematics, in computer science, but there are no questions here on cryptography."

But what the release changed for the crypto industry is the rate of potential change for key pieces of infrastructure, like wallets. Problems that resisted people for generations fell in batches, at a few hours of compute apiece.

Drake's post put it plainly: "Recent days have been humbling for human mathematical intuition. Long-held, unquestioned hypotheses have fallen." The assumption guarding your wallet is one of those hypotheses. It just hasn't been asked yet.

What now?

In light of the latest OpenAI findings, Drake's recommendation for the best security was specific.

Large holders should "calmly begin planning" for what he called bunker mode: move most of your funds to a brand-new address that has never sent a transaction, so your public key stays hidden behind the scramble, and move whatever's left to another fresh one.

He said it was "now reasonable to brace for the possibility that ECDSA breaks before qday," meaning before quantum computers arrive, "in the worst case in months not years." He also wrote "Don't rush," and said the Ethereum roadmap's timelines "must now be revisited and accelerated." (This newsletter covered that roadmap, Lean Ethereum, in July, when quantum-safe signatures were already marked urgent.)

Vitalik Buterin agreed the direction is real and that ECDSA "might fall even faster than expected." He also declined to tell anyone to scramble, and gave the best line of the week, per Unchained: "I personally have lost more money in botched migrations than I have lost in all hacks combined."

Others were blunter. Yehuda Lindell, who leads cryptography at Coinbase, called the post "a really bad take," with "no evidence whatsoever pointing to a break of decades old hardness assumptions like elliptic curve cryptography."

Ledger's CTO Charles Guillemet said avoiding address reuse was "reasonable for large holders" and warned about "operational mistakes that lose funds with higher probability than the scenario being mitigated."

Dankrad Feist made the darkest point: if exposed keys ever become crackable, the market goes to zero and your hidden key won't save its price.

The same day, Europol's cybercrime centre published Quantum Computing and Cryptocurrencies, written with a different attacker in mind and landing in the same place.

"Cryptocurrencies will not collapse due to quantum computing," it says, as quoted by Decrypt. Wallets are "the primary point of exposure," and for keys already exposed onchain, "the only solution is pre-emptive migration."

Addresses already in the open

Glassnode estimated in May that 6.04 million bitcoin, about 30 percent of everything issued, sat at addresses whose public key was already visible, per Decrypt.

Project Eleven, a firm that sells post-quantum security, tracks more than 14 million funded Bitcoin addresses with exposed keys, according to The Block.

Drake counted around 20,000 early addresses linked to Satoshi Nakamoto, each holding 50 bitcoin, with keys that have been public since the network's first year and owners who haven't touched them in more than fifteen.

Drake's own scenario turns that into a grim kind of insurance. If cracking a key took something like a week of heavy computing, an attacker would start with the biggest prizes, and the biggest prizes are Satoshi's. He calls it "Satoshi's shield." A small holder's best protection is a famous fortune standing in the line ahead of them.

Even the coins that can move face a traffic problem. The Europol report cites a 2024 estimate that migrating every Bitcoin output would take at least 76 days of the network doing nothing else.

Ethereum's version is worse in one way, because every account that has ever sent anything has shown its key. Jacob Creech argued that Solana users don't need bunker mode at all, because their keys come from a hashed seed that never appears onchain.

Changing the locks

Here the gated systems have a real advantage, which is an important distinction to make since this newsletter advocates for open rails.

Tether can freeze any USDT it issued. Conduit, a cross-border payments firm, sued Tether in federal court in Manhattan this week over $2.76 million of USDT that's been frozen for more than a year, calling the wallet the "equivalent of [its] operating bank account."

That's the ugly side of the switch. The useful side is that an issuer who can freeze a token can also reissue it on a new lock, and the holder never has to understand why.

The regulated version of that power is being written right now. On September 24 the Federal Reserve proposed rules under the GENIUS Act requiring Board-supervised stablecoin issuers to "fully back their stablecoins" with Treasury bills and similar assets, and setting up an application path for banks that want to issue them.

In Europe, the markets regulator ESMA told licensed platforms on October 8 to stop serving stablecoins without EU authorization and to clear out remaining client holdings within three months.

The token may run on a public rail while somebody accountable still holds the master key, can halt it, and can swap the signature scheme on a Tuesday.

For a bank deciding where to hold tokenized cash, that's a feature. A key-rotation plan with a phone number attached is easier to put in front of a risk committee than a governance forum.

Open systems don't get that option. Their upgrade path is rough consensus, wallet updates, and millions of people deciding to move, which is exactly the process Buterin says costs him more than hackers do.

The same machine can check the upgrade

The release cuts the other way too, and that's the part I'd keep an eye on.

About 300 of OpenAI's results came with proofs a computer checked line by line. Lean doesn't care who wrote the proof or how famous they are. It either compiles or it doesn't.

The same tool works on code. A smart contract is a set of promises about money ("this vault never pays out more than it holds"), and those promises can be written as theorems and checked the same way.

Until now, writing those proofs was slow, specialized work that only the largest protocols paid for. A model that can produce hundreds of machine-checked math proofs in a week makes that work much cheaper.

Boneh pointed at a quieter use: aim the models at a security assumption for a month, and if they can't break it, you've earned more confidence than decades of human failure gave you. He also noted that some internet connections already lock data with elliptic curves and lattice math at the same time, so breaking one isn't enough. That habit of building in more than one lock, which cryptographers call crypto agility, is the thing open money most needs as a new frontier.

There's a catch Buterin flagged himself. The lattice schemes lined up to replace ECDSA, like the NIST standard ML-DSA, also rest on unproven assumptions, and he named them as a new area of risk. Moving to the new lock buys time, and only time.

So the honest version of open money's pitch now has two halves. Anyone can check the rules, and anyone can check the math, with software. This week made that system of checks and balances cheaper and more urgent at the same time.

The honest case against this read

The strongest objection is the obvious one: nothing broke. Boneh, Lindell and the a16z panel all said the release taught us nothing new about elliptic curves, and they're the right people to ask.

If I'm writing about a risk with no demonstrated attack, I might just be amplifying a mood.

What keeps me on this side is that the argument doesn't need an attack. It needs only that the timing became harder to estimate, and a researcher at the Ethereum Foundation, a European police agency and the person who co-founded Ethereum all moved their estimates in the same week.

The second objection is Buterin's: the cure can hurt more than the disease. Mass migrations cost money through typos, lost keys and the fake "migration" services that tend to follow headlines like this week's.

I think that's the best reason to plan now, while there's no deadline, rather than in a panic later. It's also a reason small holders should probably do nothing yet.

And OpenAI hasn't released the model or the prompts, so mathematicians like MIT's Andrew Sutherland say single-agent claims should be treated as unverified "until and unless they release the model and people can replicate their results."

Tokenized Treasuries Don’t Trade: Pantera’s Data, BlackRock’s BUIDL and the SEC Innovation Exemption (2026)

Pantera’s State of Tokenization found that whitelisted tokens hold 59% of tokenized value and produce 0.2% of onchain trading. BlackRock’s $2.2 billion BUIDL fund has 104 holders.

Daniel McGlynnDaniel McGlynn

Sources

[1] OpenAI. "math" repository, README and CONTENTS. 722 manuscripts in 372 result families, approximately 4,000 problems posed, average of three hours of ChatGPT Pro thinking compute per result, verification caveats. https://github.com/openai/math/blob/main/README.md

[2] Implicator. "OpenAI posts 372 AI math results, then withdraws three papers over a sign error," October 8, 2026. Lean coverage of 300 of 719 top-line results as of October 7, withdrawals, spokesperson statements, Andrew Sutherland quote. https://www.implicator.ai/openai-posts-372-ai-math-results-then-withdraws-three-papers-over-a-sign-error/

[3] Math Scholar. "OpenAI stuns mathematicians with 722 new papers," October 2026. Example results and Martin Bridson quote. https://mathscholar.org/2026/10/openai-stuns-mathematicians-with-722-new-papers/

[4] a16z crypto. "Can AI really break cryptography? Experts weigh in," October 10, 2026. Dan Boneh, Eddy Lazzarin and Justin Thaler. https://a16zcrypto.substack.com/p/can-ai-really-break-cryptography

[5] Justin Drake, post on X, October 7, 2026. https://x.com/drakefjustin/status/2107837081313505768. Quotations as reported by FXStreet, Unchained and The Block.

[6] The Block. "Ethereum researcher Justin Drake calls for 'bunker mode' planning over private key recovery risk," October 7, 2026. Satoshi-linked addresses, "Satoshi's shield," Project Eleven address count. https://www.theblock.co/news/ecosystems/2026-10-07-ethereum-researcher-justin-drake-bunker-mode-planning-private-key-recovery-risk-417954

[7] Unchained. "Ethereum Researcher's 'Bunker Mode' Call Sparks Debate Over Whether AI Math Threatens Crypto Keys," October 2026. Quotations from Vitalik Buterin, Yehuda Lindell, Charles Guillemet, Dankrad Feist. https://unchainedcrypto.com/ethereum-researchers-bunker-mode-call-sparks-debate-over-whether-ai-math-threatens-crypto-keys/

[8] FXStreet. "Cryptocurrencies face new security risk as Ethereum researcher warns of potential ECDSA break," October 8, 2026. https://www.fxstreet.com/cryptocurrencies/news/cryptocurrencies-face-new-security-risk-as-ethereum-researcher-warns-of-potential-ecdsa-break-202610080101

[9] Europol, European Cybercrime Centre. "Quantum Computing and Cryptocurrencies," October 7, 2026. https://www.europol.europa.eu/publications-events/publications/quantum-computing-and-cryptocurrencies. Quotations and the Glassnode exposed-supply estimate as reported by Decrypt, "Europol warns crypto wallets are primary risk for quantum attacks." https://decrypt.co/380268/europol-warns-crypto-wallets-are-primary-risk-for-quantum-attacks

[10] The Block. "Conduit sues Tether over $2.76 million freeze," October 6, 2026. https://www.theblock.co/news/regulation/2026-10-06-conduit-sues-tether-usdt-freeze-417831

[11] Board of Governors of the Federal Reserve System. Press release on GENIUS Act payment stablecoin proposals, September 24, 2026. https://federalreserve.gov/newsevents/pressreleases/bcreg20260924a.htm

[12] European Securities and Markets Authority. "ESMA sets out supervisory expectations on services related to unauthorised stablecoins," October 8, 2026. https://www.esma.europa.eu/node/235756

[13] European Central Bank, Isabel Schnabel. "Central banks on-chain," The Future of Money Conference, Bank of England, October 1, 2026. Pontes launch date and settlement model. https://www.ecb.europa.eu/press/key/date/2026/html/ecb.sp261001_1~a0be67193b.en.pdf