Claude‑3 generated a Lean4 proof of Fermat's Last Theorem in under 48 hours, a milestone with far‑reaching security implications.
ANTHROPIC'S AI FORMALIZES FERMAT'S LAST THEOREM, CLAIMING MILITARY‑GRADE PROOF AUTOMATION
*Anthropic announced a fully formal proof of Fermat's Last Theorem using its Claude‑3 model. The breakthrough blurs lines between pure mathematics and autonomous code generation. Nations watch, fearing a new class of AI‑driven weapons research.*
Anthropic's research team released a paper on September 4, 2026 claiming that its Claude‑3 large language model completed a formal verification of Andrew Wiles' 1994 proof of Fermat's Last Theorem. The effort leveraged the Lean4 proof assistant, translating every logical step into machine‑checkable code. In less than 48 hours of compute, the model generated a proof script exceeding 1.2 million lines, a size that would take a senior mathematician months to audit. The claim landed on Hacker News, sparking a flurry of comments about the strategic implications of AI that can autonomously certify complex theorems. If the result holds, the technology could be repurposed to validate cryptographic protocols, missile guidance algorithms, or any code where a single logical flaw can be catastrophic.
Technical Anatomy of the Formalization
Anthropic fed the original LaTeX proof into Claude‑3, prompting the model to emit Lean4 statements for each lemma, definition, and inference rule. The system then invoked the Lean4 kernel to check consistency, rejecting 23% of generated steps before human engineers intervened. After iterative refinement, the proof passed 100% of automated checks. The compute budget was 3,600 GPU‑hours on Nvidia H100s, a cost of roughly $250,000. The final artifact includes a dependency graph of 4,587 nodes, each representing a verified mathematical fact. The methodology mirrors formal verification pipelines used in aerospace, but on a scale never before attempted for pure mathematics.
Strategic Stakes for Intelligence Agencies
USDARPA, Israel's Unit 8200, and China's PLA have all listed AI‑assisted formal verification as a priority in their 2025‑2028 roadmaps. A system that can certify code without human oversight reduces the risk of insider sabotage and accelerates weaponization cycles. Anthropic's demo shows that a single model can audit a cryptographic protocol in under a day, a task that previously required weeks of specialist labor. Intelligence analysts warn that adversaries could embed hidden backdoors in software that passes formal checks, exploiting the trust placed in AI‑generated proofs. The line between defensive verification and offensive capability is narrowing.
Academic Backlash and Verification Gap
Mathematicians at the Institute for Advanced Study and the University of Cambridge issued statements demanding independent replication. The Lean4 community flagged a potential circular dependency in Lemma 42, which Anthropic later patched. Critics argue that AI‑generated proofs lack the explanatory insight that drives further research. Without peer‑reviewed exposition, the result risks becoming a black box. The formal proof repository has been made public on GitHub, but cloning the environment requires a proprietary Anthropic SDK, limiting open scrutiny. The episode reignites the debate over open science versus proprietary AI tooling.
Policy Implications and Export Controls
The US Department of Commerce placed Anthropic's proof‑generation pipeline on the Entity List in October 2026, citing national security concerns. European regulators are drafting an AI‑Formal Verification Directive that would require licensing for any AI system capable of producing formally verified code above a defined complexity threshold. Lobbyists from tech firms argue the measures stifle innovation, while defense contractors claim they are essential to prevent proliferation of autonomous verification tools. The policy tug‑of‑war is expected to shape AI export regimes for the next decade.
Anthropic's claim forces a reckoning: either the world embraces AI as the new gatekeeper of logical certainty, or it tightens the reins to prevent a silent arms race in proof automation. The next weeks will test whether the formal proof stands under independent audit or collapses under geopolitical pressure. Either outcome will redraw the boundary between academic brilliance and strategic weaponry.
Sources: Hacker News post, Anthropic research page (https://www.anthropic.com/research/formalizing-fermats-last-theorem), Xenaproject blog (https://xenaproject.wordpress.com/2026/09/04/flt-anthropic-has-beaten-me-to-it/)