Whether AI is weakening the cryptography we rely on: attacks and cryptanalysis found or sped up by AI, AI mathematics with a bearing on lattices, elliptic curves, hashes and number theory, parameters revised and schemes dropped because of it, and what cryptographers and protocol designers say about the threat. Topics and rules ↓
Lavit, Seinturier, and Véron show a transformer-based model identifies Goppa codes from random linear codes by exploiting low-weight-row artifacts and combinatorial properties, not hidden algebraic structure. The finding suggests AI can extract general code properties relevant to code-based cryptography security even after removing training artifacts.
Anthropic showed Claude reconstructed the public seed behind P-256, a NIST elliptic curve used for Ethereum passkey signing, by initializing an old C-library generator with timestamp Aug. 16, 1997, 01:52:34 UTC and advancing to candidate 1,842. This resolves longstanding concerns about whether NSA-supplied curve seeds contained deliberate backdoors.
Researchers released LeanPRM, a dataset of 152,000 labeled proof steps across 10,634 formal Lean proofs, generated automatically without human annotation. This removes the bottleneck for training AI verifiers on mathematical proofs, enabling scalable step-level verification of AI-generated proofs.
Researchers released LeanPRM, a dataset of 152,000 labeled proof steps across 10,634 formal Lean proofs, generated automatically without human annotation by pairing natural-language proofs with formal Lean code and using Lean's checker to assign correctness labels. A Qwen3-4B verifier trained on this data reached 74.7% exact step-label accuracy on held-out benchmarks, compared to 60.3% for models trained on the existing PRM800K dataset. This removes the human-annotation bottleneck that previously capped training-set size, making it practical to build scalable AI verifiers for mathematical proofs at a time when step-level verification of AI math is actively disputed.