Vitalik Buterin on AI risk to cryptography: is the two-year claim accurate, and what does he advise?

The recommendations in the circulating summaries match his original X post (7 Oct 2026). The "two years" claim is a personal forecast, not an established fact. It has partial support from one reported lattice result and no proof for ML-DSA or ECDSA.

What he actually said

He does not recommend anyone scramble to move funds today. He says AI-accelerated math risks should be taken seriously, and exposure should be minimised to quantum-vulnerable and potentially AI-vulnerable cryptography.
He calls ML-DSA, FHE and lattices the core new area of risk. He says there is "a good chance" lattice security takes serious hits in the next two years of AI math. That is a probability judgment, not a prediction of a full break.
ECDSA "might fall even faster than expected", which is why he backs the "fresh address" advice.
His threat model: factoring went from a naive 2^(n/2) to number field sieve speeds after decades of work. Similar hidden structure could exist for elliptic curves and lattices, and AI might find it.
Hash-based schemes are preferred wherever possible, because hashes are designed to have no algebraic structure. He says Lean Ethereum has gone "hash-only" for a year, with WOTS or SPHINCS-style signatures and no lattices, ML-DSA or Falcon. There is no reason yet to pad hash sizes.
Public-key encryption cannot be done with hashes alone. For anything with structure he suggests roughly 10x larger keys.
Cointelegraph quotes him: "If it's not difficult for you, keeping your funds in addresses which have not yet been used to make a transaction is a good idea." He also said: "I personally have lost more money in botched migrations than I have lost in all hacks combined."
Cointelegraph's news feed also reported he suggested multisig signers ideally switch to a new key after every transaction, with the caveat "don't rush anything". The Wu Blockchain summary adds off-chain multisig signing "where practical".

Is the claim accurate?

Support, as reported by X posts: Anthropic's AI model reportedly found a flaw in HAWK, a lattice signature candidate in NIST's process. NaoX says this took about 60 hours and that HAWK was withdrawn on 29 July. BlockClaimed says NIST disclosed it and that the flaw does not affect finalized ML-KEM or ML-DSA. I found no news-wire coverage of HAWK, so this rests on social posts.
Justin Drake, an Ethereum Foundation researcher, goes further in his post. He says ECDSA could break before quantum computers arrive, "in the worst case in months not years", citing recent AI-driven math results. This is his opinion and is more alarmist than Vitalik's.
Against: no break of ML-DSA or ECDSA has been demonstrated. Vitalik himself says this is a "very plausible world", not a certainty. Some X reactions, such as this one, note that Vitalik's tone is calmer than Drake's.

Practical takeaways

Both Vitalik and Drake say not to panic or rush. Drake specifically warns a rushed migration would do more harm than good.
If it is easy, hold funds on addresses that have never sent a transaction, so the public key stays hidden behind a hash. Once you sign from an address, move the remaining funds to a new one.
Drake says this can be a fresh address from the same seed phrase.

Vitalik's post · Drake's post · Cointelegraph · Wu Blockchain