When AI Breaks Math: The Cryptographic Alarm Hidden in the Jacobian Counterexample
Hook
Over the past week, two independent AI models—Anthropic’s Claude Fable and OpenAI’s Codex—found counterexamples to the 3D Jacobian conjecture. This isn’t a headline about machine learning benchmarks. It’s a cryptographic alarm. The Jacobian conjecture, unsolved since 1939, asks whether a polynomial map with a non-zero Jacobian determinant is always invertible. The AI found maps that are not invertible, disproving the conjecture in three dimensions. The math world is buzzing. But for anyone who works with zero-knowledge proofs or public-key cryptography, the subtext is chilling: if AI can navigate the vast search space of polynomial maps to find a counterexample to a 90-year-old problem, what prevents it from finding a practical attack on RSA or elliptic curve cryptography?
Context
First, the basics. The Jacobian conjecture is a problem in algebraic geometry. It asks: if you have a polynomial function from n-dimensional space to itself, and its Jacobian determinant is a non-zero constant everywhere, must the function be globally invertible? For n=1 and n=2, the answer is yes. For n=3, mathematicians suspected it was true. Claude Fable and Codex each independently generated polynomial maps where the Jacobian determinant is constant non-zero, but the map is not one-to-one. These are valid counterexamples.
Why should a blockchain researcher care? Because the security of most crypto systems rests on mathematical assumptions that are equally vulnerable to novel counterexamples. RSA relies on the difficulty of factoring large integers. ECC relies on the discrete logarithm problem. Both are assumptions, not proven facts. The Jacobian counterexample shows that AI can now systematically explore mathematical spaces where humans have only skimmed the surface. The same pattern-matching and combinatorial optimization could be applied to finding factorisation algorithms or discrete log shortcuts.
Core: The Mathematics of Vulnerability
Let’s get technical. The AI didn’t invent new mathematics. It combined existing concepts—polynomial maps, Jacobian matrices, irreversibility—in novel ways. According to the reports, Claude Fable was guided by mathematician Levent Alpöge, who prompted the model to generate candidate maps and verify their Jacobian determinants. The model iterated, producing thousands of candidates until it found one that satisfied the conditions. Codex, a code-generation model, did the same without internet access, relying on its training data.
The critical insight: these models are not proving theorems; they are searching a manifold of possible mathematical structures. The search space for polynomial maps in 3 dimensions with integer coefficients up to a certain degree is astronomically large. Human mathematicians cannot brute-force it. AI can, using learned heuristics to focus on promising regions. This is fundamentally different from traditional computer algebra systems, which apply deterministic algorithms. AI brings approximate reasoning to a domain that previously required exact proof.
In my own work auditing zero-knowledge circuits, I’ve seen how small breaks in mathematical assumptions cascade. In 2022, I built a minimal zkSNARK generator in Rust. The Groth16 proving system relies on the hardness of the discrete logarithm in certain elliptic curve groups. A single counterexample to that assumption would render every zkSNARK built on that curve insecure. The Jacobian counterexample is not directly about discrete logs, but it demonstrates the pattern: AI can find holes in well-studied mathematical conjectures.
Now, translate this to cryptography. The security of Schnorr signatures, ECDSA, and many ZK proving systems (like PLONK and Halo2) depends on the discrete logarithm assumption. If an AI were to find a polynomial-time algorithm for discrete log on standard curves like secp256k1, the entire cryptocurrency ecosystem would collapse within hours. The exact same technique used for the Jacobian—searching for counterexamples in a parameter space—could be applied to finding “weak” curves where the discrete log problem is easier than expected.
But the threat is more subtle. Even without a complete break, AI could find specific instances where the mathematical assumptions fail for particular inputs. For example, it might find a set of elliptic curve points where the discrete log can be computed efficiently using a novel algebraic manipulation. This would allow targeted attacks: compromising one wallet, forging one signature, without breaking the entire system. The Jacobian counterexample was exactly that—a specific, unexpected failure of a general statement.
Privacy is a feature, not a bug. But privacy based on weak math is a feature that can be exploited. Many privacy-preserving protocols, including zk-SNARKs and zk-Rollups, rely on assumptions about the hardness of polynomial equations. The Jacobian case is a polynomial equation problem itself. The fact that AI can find exceptions to polynomial invertibility should make every ZK engineer revisit their trusted setup assumptions.
I recently analyzed the security of a multi-party computation protocol for a DeFi lending platform. The protocol used a NIZK proof based on the security of the BLS12-381 curve. The security proof assumed that the discrete log in that curve’s pairing groups is intractable. If AI finds a counterexample to that assumption—say, a specific pairing that leaks the exponent—the entire privacy layer becomes transparent. The Jacobian counterexample is a proof-of-concept that AI can find such needle-in-a-haystack vulnerabilities.
Contrarian: The Blind Spot Is Verification, Not Discovery
The common reaction is fear: “AI will break all crypto.” But the real blind spot is different. The Jacobian counterexample was validated by human mathematicians. They checked the map, confirmed it was a valid counterexample. The AI didn’t need to prove it; it just generated candidates. The bottleneck today is not AI finding counterexamples—it’s humans verifying them. In cryptography, the same applies: AI might generate a million candidate attacks on RSA, but verifying each one requires running a breaking algorithm. The real crisis will come when AI can not only find but also formally verify its own counterexamples, closing the loop without human involvement.
Math doesn’t negotiate. The Jacobian conjecture is either true or false. AI found it false. But the implications for crypto depend on whether the same capability extends to problems like factoring or discrete log. I argue the probability is high, but the timeline is uncertain. What we can control is our response: we must shift from trusting mathematical assumptions to verifying cryptographic implementations continuously. Every circuit, every curve, every proof system should be subject to automated, AI-assisted auditing.
Takeaway
The AI breakthrough in the Jacobian conjecture is a canary in the coal mine. It proves that AI can navigate abstract mathematical spaces and find results that eluded human experts for decades. The same tools will soon be applied to cryptographic primitives. We have a narrow window—perhaps 3 to 5 years—to migrate to post-quantum and AI-resistant schemes. Zero-knowledge proofs themselves must be proven secure not just against classical attackers, but against AI-powered cryptanalysis. Code is law, but bugs are reality. The next bug might be in the mathematics itself.