vitalik: ai math, not quantum, may crack crypto's locks first

daily ai × crypto news in 5 minutes. no hype, just what matters.

evening. a break from hacks at the top, almost. devcon 8 opened in mumbai this week, and vitalik buterin used it to move the goalposts on crypto security: the thing to fear may not be a quantum computer, but ai doing decades of math in a few years. and the lock most likely to crack first is the "quantum-safe" one. also tonight: anthropic shows a zero-knowledge proof of which ai model answered you, a docs host used by half of crypto got hit by an "ai-agent" attack, and a french bank is named in a $133m+ crypto theft.

the new worry. on nov 3, the day he opened devcon 8, vitalik buterin wrote on x that his fear is that shortcuts for breaking today's crypto math "exist, and we've just not yet managed to find them – but ai speedrunning 50+ years of math in 5 years will." he named the math that protects private keys on bitcoin and ethereum (the elliptic curve discrete log problem), the pairings behind ethereum validator signatures, and lattices. (beincrypto, quoting vitalik)

the three locks. in a longer post on oct 7 he spelled it out: most people think "elliptic curves broken, hashes safe, lattices safe", meaning quantum computers kill today's wallet keys while lattice math and hash functions survive. lattices are the basis of ml-dsa and other leading post-quantum standards, the migration everyone is planning. his view: "there is a good chance that the concrete security of lattices will take serious hits from the next two years of ai math." (vitalik on farcaster)

why it's not crazy. it has happened before. factoring big numbers naively takes 2^(n/2) time, but decades of better algorithms cut that down so far that rsa keys need to be ~400 bytes instead of 64. "what if there are skeletons in the closet like that, both for elliptic curves and lattices, that we are simply not smart enough to discover - but bots soon will be?" his practical rule for anything built on lattice cryptography that must last: "multiply the key sizes by 10." (vitalik on farcaster)

ethereum's bet: plain hashes. for the past year, he says, ethereum's "lean" roadmap has gone hash-only: "no lattices, no ml-dsa, no falcon", with signatures built from hash-based schemes. on the same opening day, ethereum foundation researcher justin drake said: "our hope is that by december 2028 we'll be able to wrap up the full post-quantum and post-ai transition", which means three more network upgrades in about 24 months. note the word "hope": the foundation's official target is still dec 2029 (set sept 7), fixed until a jan 2027 review with outside quantum experts. (coinmarketcap, cointelegraph, ethereum foundation)

what to do (and not do). vitalik: "i don't recommend anyone scramble to move their funds to new wallets today." but he adds that ecdsa, the signature behind most wallets, "might fall even faster than expected, hence the 'fresh address' recommendation." a fresh address is one that has never sent a transaction, so its public key hasn't been revealed on-chain yet. on stage he added a hardware angle: "not your silicon, not your life. we need more open and more verifiable silicon," as a hypothetical risk, naming no manufacturer. (farcaster, the crypto times)

why it matters. this is a risk view, not a found exploit. nobody has shown an ai breaking elliptic curves or lattices. but it flips the usual story: the industry has treated "post-quantum" as "safe", and vitalik is saying the replacement could be the weak spot, because ai pressure hits any math with hidden structure. it's the same thread as #017, where an ai scan found real bitcoin core bugs, and as #006, where weak wallet math, not a broken chain, drained 1,404 wallets. more in ethereum.

anthropic open-sources a zero-knowledge proof of which ai model answered you. in a research post (nov 2), anthropic's frontier red team released flashprove: the provider publishes a fingerprint ("commitment") of its model weights, then attaches a zero-knowledge proof to replies showing that those exact committed weights produced those tokens, without revealing the weights. it targets routing services or attackers quietly swapping in a cheaper or tampered model. (anthropic, techcrunch)

the numbers: on llama 4 maverick, a 400b-parameter model (~232 gb quantized), one h100 proves about 110 tokens per second, each proof is ~19 mb, and checking it takes ~17 seconds on a 24-thread cpu. the cost is ~70–95× the gpu time of normal inference, and it only covers integer, greedy decoding up to 1,024 tokens, not normal sampled chat. techcrunch says there are no plans to deploy it at anthropic or any other lab; it's a research prototype on github. and note: it proves a reply matches a commitment, not that the commitment is "claude"; you still have to trust who published it. the crypto link: it's the same zk tech that secures ethereum rollups.

gitbook, the docs host for much of crypto, hit by an "ai-agent" attack. gitbook says it was the target of "a sophisticated, highly targeted attack using ai agents": at 19:00 cet on nov 2 the attacker got some of its infrastructure credentials. gitbook took the app and authenticated docs offline, replaced every credential and invalidated every customer's api tokens. possibly accessed: integration tokens, git sync secrets, api tokens and questions visitors asked gitbook's ai assistant. it found no evidence content was modified and called it resolved on nov 4, 11:36 cet. (gitbook)

pancakeswap pulled its docs as a precaution, reminding users that "docs never ask you to connect your wallet or approve tokens", then moved them to its own servers. gitbook hasn't said which ai was used or how the credentials leaked. the real lesson is quieter: a docs page is where users copy contract links and install commands, so a tampered one is phishing with the project's own logo. if you use gitbook's mcp server, re-authenticate. (decrypt)

zachxbt names french bank crédit mutuel arkéa in a $133m+ crypto theft. the on-chain sleuth first flagged an unnamed victim losing more than $106m from exchange accounts on monday, then found $27m more and named the bank on wednesday, saying the total could be higher with otc trades. bitquery traced $108.8m in usdc withdrawn from kraken, crypto.com and okx over 21 hours on nov 1–2, much of it swapped into ether, which no issuer can freeze. (coindesk)

the bank confirmed to coindesk a cyberattack that "enabled fraudulent transactions", says it's under control and clients face no losses, but did not confirm the amount or that crypto was involved. crypto.com says it blocked linked funds. zachxbt has "a high degree of confidence" it was not north korea's lazarus group, unlike the binance hack in #021. (coindesk)

sparrow wallet: a desktop bitcoin wallet built for security and privacy, with full coin control, labels on every transaction and coin, multisig, built-in tor, and support for common hardware wallets (usb and airgapped). it shows exactly which addresses and coins you hold, which makes it easy to avoid reusing addresses. free and open source (apache-2.0, latest release 2.5.6 on oct 29). for bitcoin holders ready to move past exchange apps toward self-custody. more picks in the gweiwire tools catalog.

whether ethereum's client teams sign up to drake's dec 2028 "mission", and whether anyone publishes an actual ai-found speedup against lattices or elliptic curves (that would be the real news). gitbook's full postmortem, promised within 24 hours. and whether crédit mutuel arkéa confirms the crypto amount.

that's the wire. the lock to worry about might be the new one. back tomorrow. — gweiwire

forward this to the friend who thinks "quantum-safe" means safe forever. new here? subscribe at gweiwire.com. not financial advice.