AI could change the cryptography risk equation faster than expected. The warning isn’t a call to panic or rush funds into new wallets, but it does put lattice-based cryptography under uncomfortable scrutiny.
The concern is that AI-accelerated mathematical research could expose weaknesses in systems many people currently consider secure. That includes ML-DSA, fully homomorphic encryption (FHE), and lattice-based cryptography. The argument isn’t that these systems have already been broken. It’s that nobody should assume today’s security estimates will remain untouched, per Vitalik Buterin’s X post.
AI Cryptography Risks Go Beyond Quantum Threats
For years, the conventional view has been simple: elliptic curves may be vulnerable to quantum computers, while hashes and lattices appear safer. But that confidence could be premature.
Factoring offers a useful comparison. Straightforward approaches scale poorly, yet decades of mathematical advances led to the General Number Field Sieve, dramatically improving the attack strategy. The concern is whether AI could uncover similarly powerful techniques against elliptic curves or lattices.
Why Lattice Security Could Face Tough Questions
The warning suggests lattice-based systems may need substantially larger parameters to preserve comparable security if new attacks emerge. One proposed scenario is multiplying key sizes by ten, though that’s a personal prediction, not an established requirement.
In that world, hash-based signatures could become more efficient where they’re practical. Ethereum’s lean roadmap has already favored hash-based approaches, including WOTS and SPHINCS+, while avoiding lattice-based signatures and commitments in zero-knowledge proofs.
Wallet Safety Still Doesn’t Justify Panic
The argument extends beyond crypto wallets to secure messaging, VPNs, Tor, and other privacy tools. Public-key encryption remains a particularly difficult problem because hashes alone cannot provide the necessary trapdoor structure.
For everyday users, the advice is measured: unused addresses may reduce exposure if elliptic-curve signatures become vulnerable sooner than expected, but risky migrations can cost money too. The AI cryptography debate is a reason to assess exposure carefully, not scramble blindly. For lattice-based systems, the key question is whether their security assumptions will withstand the next wave of AI-driven mathematics.