Can AI Break Cryptography? What the Experts Actually Think

The mathematical and computer-science advances that triggered concern have not demonstrated a break in cryptography: none of roughly 700 published problems addressed cryptographic systems, and the labs’ own tests and results remain unknown. Dan Buene and Eddie Lazarin said the reaction reflected broad progress, not evidence that a deployed system had been compromised. The security of systems in use still rests on assumptions such as the difficulty of discrete logarithms and lattice problems; after decades of attempts, no efficient attacks against those assumptions have been found. Buene added that the advances do not come close to resolving P versus NP.

AI could find an attack human researchers have missed, but that is not a new category of risk: a person could also discover one. The deeper uncertainty is whether the belief that no efficient attack exists is wrong, rather than whether an already-known algorithm has simply escaped notice. The speakers said there is no new evidence that elliptic curves or another specific cryptographic assumption has become less secure, and the latest models they discussed could not break the post-quantum candidates under consideration.

Buene urged frontier AI labs to disclose what they have tested against basic cryptographic problems and standardized, deployed systems. A record of failed attempts could increase confidence, though only until a substantial model upgrade. Even a hypothetical attack that reduced a practical system’s security by 10 bits would be valuable evidence. One speaker suggested a month of model testing might build confidence faster than a decade of conventional scrutiny; another proposed that labs might spend $20 million or $50 million testing a scheme with a highly capable model. Such testing could inform confidence, but cannot prove a system secure.

The possible AI risk is separate from the longstanding quantum-computing threat, though both could affect elliptic-curve cryptography. A practical break of elliptic-curve discrete logarithms could endanger banking, health care, government communications, and other digital infrastructure, not only blockchains. Justin Drake warned of a possible complete break in the coming months and advised large Ethereum cold wallets to consider “bunker mode” or avoid exposing public keys on-chain. Vitalik Buterin offered a more cautious speculation that AI might find better attacks, and noted that botched migrations had cost him more personally than all hacks. The speakers said much of the proposed bunker-mode guidance amounts to existing institutional security hygiene; Coinbase said it already follows many of Drake’s specified measures.

NIST is standardizing alternatives including lattice-, code-, multivariate-polynomial-, isogeny-, and hash-based systems. The speakers emphasized crypto agility—the ability to change designs if an assumption fails—as an established defense, and said some mitigations for elliptic-curve systems are already prepared or deployed because of the quantum threat. But “post-quantum” does not mean “post-AI.” In blockchain, hash-based signatures can build on hash functions already used in systems such as Merkle trees and may be easier to adapt if a hash function fails. Bitcoin Core developers and people in the Ethereum Foundation space are exploring ways to remove elliptic-curve assumptions in favor of hash functions. Hash-based signatures, however, are larger, stateful, harder to use, and less suited to advanced applications; lattice-based systems support features such as threshold signatures and non-hardened key derivation.

Hash-only cryptography cannot provide good encryption key exchange, so it cannot replace every cryptographic system across digital infrastructure. Ethereum’s hash-only approach also uses recursive SNARKs, whose nested proof verification is complicated and creates a large bug surface; the speakers said the security of SNARK recursion is not well understood, even though AI tools may help find bugs. For ordinary users, their advice was not to panic or change anything absent new information about standard cryptography. A full move away from elliptic curves within a few years is unrealistic, though some internet connections already use hybrid exchanges combining elliptic curves and lattice-based systems, including X25519 ML-KEM-768. If both elliptic-curve and lattice-based systems were broken, a third system would require introduction and standardization, taking longer. The speakers urged institutions to follow existing security practices, startups to assess software assumptions and constrain AI agents with boundaries and permissions, and anyone discovering a break to disclose it responsibly so industry and government can coordinate updates.