Over the past week, two AI models—Anthropic's Claude Fable and OpenAI's Codex—independently found counterexamples to the three-dimensional Jacobian Conjecture, a problem unsolved since 1939. The code whispers what the auditors ignore: these models aren't just writing Solidity anymore; they are discovering new mathematical structures. For those of us in DeFi security, this isn't a physics paper—it's a canary in the cryptography coal mine. The same pattern-recognition that found a polynomial map with a non-invertible Jacobian could, in a few iterations, find an attack on the discrete logarithm problem underlying every Bitcoin address, every Ethereum wallet, every cross-chain bridge.
Context: The Math Behind the Chain
Blockchain security rests on a handful of mathematical assumptions. The security of ECDSA (secp256k1), Ed25519, and BLS signatures all depend on the hardness of the discrete logarithm problem (DLP). While quantum computers remain a theoretical threat, AI has now demonstrated a practical ability to navigate high-dimensional search spaces where human intuition fails. The Jacobian Conjecture counterexample is a proof of concept: AI can combine known concepts (invertible maps, polynomial rings) in unexpected ways to produce a counterexample. The step from “find a non-invertible map” to “find a way to compute a private key from a public key” is not trivial, but the neural architecture that enables the former is the same that could enable the latter. In my five years auditing DeFi protocols, I’ve seen countless smart contract vulnerabilities arise from assumptions that code would behave predictably. AI now challenges the predictions at the mathematical foundation.
Core: The Code-Level Risk in Your Wallet
Let’s go granular. Every time you sign a transaction, the elliptic curve arithmetic is executed by your hardware wallet or browser extension. The security proof for ECDSA assumes that the discrete logarithm cannot be efficiently computed. But AI models like Codex have shown they can generate thousands of candidate solutions per minute, filter them through symbolic verification, and refine outputs based on error signals. I’ve run internal simulations during my audit of a major cross-chain bridge’s custody module in 2024: a hypothetical AI trained on ECC literature could attempt to learn the mapping from public key (Q = dG) to private key (d) by treating it as a high-dimensional regression problem. The model doesn’t need to “invent” a new number-theoretic algorithm; it only needs to find a counterexample to the assumption that no efficient mapping exists. The Jacobian result proves AI can already find mathematical counterexamples where human mathematicians failed. The leap from “map” to “discrete log” is a difference of domain, not of fundamental capability. Logic holds when markets collapse, but the market hasn’t priced this risk yet.
Contrarian: The False Comfort of “Post-Quantum Ready”
Many projects now advertise “post-quantum security” using lattice-based signatures like SPHINCS+ or CRYSTALS-Dilithium. The contrarian truth: these schemes are not immune to AI discovery. The LWE (Learning With Errors) problem, which underlies most post-quantum crypto, has its own class of vulnerabilities; in 2025, a team at Google DeepMind showed that a transformer can solve small LWE instances faster than lattice reduction algorithms. The bigger blind spot is the migration risk. Every wallet, every smart contract, every governance vote that switches from ECDSA to a new signature scheme introduces a window of cryptographic chaos. I audited a stablecoin issuer that froze addresses following Circle’s compliance policy; that centralized power is trivial compared to a future where an AI can compute private keys for any address on a legacy curve. The industry’s obsession with scalable consensus and MEV extraction ignores the silent, compounding risk that the cryptographic foundations themselves could be undermined not by a nation-state, but by a model trained on GitHub and arXiv. Yellow ink stains the white paper; the assumption of mathematical hardness is the paper, and the ink is spreading.
Takeaway: The Vulnerability Forecast
Over the next 3 years, every DeFi protocol with a TVL exceeding $100M should sponsor an internal red-team using frontier AI models to probe their signature libraries. The counter-example finding is not a one-off. It is a signal that AI’s capability to navigate mathematical search spaces is accelerating faster than our ability to harden cryptographic primitives. When the first AI-discovered discrete log attack emerges—and I believe it will—there will be no grace period. The hash remains, but entropy increases. If your yield-bearing vault still relies on ECDSA without a backup migration plan to a hash-based or ISogeny-based signature, you are betting against a rival that learns, iterates, and finds the gaps in mathematical proofs. The code whispers, and the auditors must now listen to a new language: the language of AI-discovered counterexamples.