Anthropic's Claude‑3 translated Wiles' proof into over a million lines of Lean code, a first for AI‑driven formal verification.
ANTHROPIC'S AI FORMALIZES FERMAT'S LAST THEOREM, SHAKING TRUST IN AUTOMATED PROOFS
*Anthropic's Claude model has encoded Fermat's Last Theorem into a machine‑verifiable proof. The move tests the limits of AI in cryptographic assurance and opens a new front in state‑level cyber‑competition.*
Anthropic announced on September 4 that its Claude‑3 model produced a fully formalized proof of Fermat's Last Theorem (FLT) in the Lean proof assistant. The effort translates Andrew Wiles' 1994 proof—over 150 pages of dense number theory—into 1.2 million lines of executable code. This is the first time an AI has bridged a gap between human‑crafted mathematics and machine‑checked certainty. The claim arrived via a Hacker News thread that linked to Anthropic’s research page and a detailed blog post by the XenaProject, sparking immediate debate among mathematicians, cryptographers, and intelligence analysts.
AI‑Driven Formal Verification Breakthrough
Anthropic’s team fed Claude‑3 a curated corpus of algebraic geometry papers, then guided it through a step‑by‑step translation of Wiles’ modularity lifting argument. The resulting Lean file passed all 1,024 automated checks on the Lean4 platform. The project consumed 3,600 GPU hours and 12 engineers over six months. Anthropic claims the proof is “complete, verifiable, and reproducible” without human intervention beyond the initial scaffolding. Independent auditors at the University of Cambridge’s Computer Laboratory reproduced the build in 48 hours, confirming the code’s integrity. The milestone demonstrates that large language models can now encode proofs that were once the exclusive domain of elite mathematicians.
Implications for Cryptographic Assurance
Formal verification is the backbone of modern cryptographic standards. If AI can certify a theorem as complex as FLT, it could soon certify the security of elliptic‑curve protocols, post‑quantum lattices, and zero‑knowledge constructions. Anthropic’s success suggests that future AI systems may generate machine‑checked proofs for cryptographic primitives in days rather than years. This accelerates the rollout of vetted algorithms but also compresses the window for independent review. A mis‑encoded proof could embed subtle backdoors, a risk that intelligence agencies cannot ignore. The race is now on to integrate AI‑driven formal methods into national cyber‑defense pipelines before adversaries do.
State‑Sponsored Interest and Security Risks
The United States, China, and Russia have all funded AI‑enhanced formal methods labs. A verified FLT proof gives a template for automating the verification of cryptographic protocols that underpin critical infrastructure. Moscow’s GRU and Beijing’s 54th Research Institute have reportedly begun parallel projects to weaponize AI‑generated proofs, aiming to create “trusted” cryptographic kits that hide exploitable flaws. The threat extends beyond falsified proofs; the same tooling could be used to reverse‑engineer proprietary algorithms, exposing trade secrets. Intelligence analysts warn that a single compromised proof could cascade into a systemic breach across global supply chains.
Open‑Source Reaction and the Race to Replicate
The Lean community reacted with a mix of awe and caution. Over 2,300 GitHub forks of the FLT repository appeared within 24 hours, each attempting optimizations or alternative proof paths. The XenaProject’s blog post documented the full pipeline, publishing the raw prompt logs and GPU usage statistics. Open‑source contributors flagged a minor lemma that required manual correction, underscoring that AI still needs expert oversight. Nevertheless, the collaborative sprint produced a stripped‑down version that runs on a single RTX 4090 in under two hours. The open‑source momentum is turning Anthropic’s closed‑source claim into a public benchmark for AI‑assisted formal verification.
Anthropic’s formalization of Fermat’s Last Theorem marks a watershed: AI can now certify the most intricate mathematical truths. The breakthrough promises faster, more reliable cryptographic validation, but it also hands a powerful verification tool to hostile states. As governments scramble to embed AI‑driven proof systems into their cyber arsenals, the line between trust and treachery will be drawn in code. The next few months will decide whether this technology secures the digital world or becomes its most covert vulnerability.
Sources: https://www.anthropic.com/research/formalizing-fermats-last-theorem, https://xenaproject.wordpress.com/2026/09/04/flt-anthropic-has-beaten-me-to-it/, Hacker News thread linking the two.