AI-GENERATED SCIENCE — THE VERIFICATION CRISIS
377 AI-generated theorems claimed — which are real? Cybersecurity news
On September 8, 2026, OpenAI reported that thousands of AI agents working together produced a machine-checked proof on the Navier-Stokes problem — one form of a Clay Millennium Prize problem — in about four days. Then on October 6, the company dropped 722 manuscripts in 372 result families from the same unreleased frontier model onto GitHub: 377 new mathematical results touching roughly 4,000 open problems. The highlights are staggering: a solution to the four-dimensional Kakeya conjecture, a zero-free region for the quasi-Riemann hypothesis beyond Re(s) > 7/8 (“If a human did this, it would be an instant Fields Medal, no questions asked” — Alex Kontorovich, Rutgers), progress on Birch-Swinnerton-Dyer, and a new 9/4 upper bound on the matrix multiplication exponent. Many results verify in Lean, which makes them all but certain to be logically correct — but some papers were already corrected or withdrawn, and mathematicians from Terence Tao (“existential crisis”) down are split between exhilaration and alarm over credit, peer review, and what happens to graduate students. This 60-second check is the verification discipline for the new era. Free, no signup.
Sep 8, 2026 — OpenAI reports AI agents produced a machine-checked Navier-Stokes proof in ~4 days.
Oct 6, 6 PM EDT — 722 manuscripts / 372 result families / 377 new results released on GitHub.
Aftermath — Many results Lean-verified; some corrected or withdrawn; OpenAI works with the Advisory Group on Mathematics and AI (Institute for Advanced Study) on release best practices; Altman declares a “new era of discovery.”
Sept 10 — OpenAI says an investigation found Tristan Buckmaster’s Codex prompts “could not have influenced the system in any way, including through training.”
Oct 13 — Scientific American (“Scientists worry AI is stealing their ideas”) reports growing scientist concern that AI companies may present users’ ideas as their models’ — citing Buckmaster’s case and a parallel Anthropic dispute with a University of Copenhagen biologist.
Why is AI proof verification cybersecurity news?
AI proof verification is cybersecurity news because the same verification gap applies to AI-generated security claims — if we cannot reliably check 377 AI-generated theorems, we cannot blindly trust AI-generated vulnerability reports either. Trust-but-verify is now a technical discipline. This 60-second check tests your verification instincts.
- Volume broke peer review. 722 manuscripts at once. Mathematicians say it will take months just to parse them — and to tell novel ideas from mash-ups of existing techniques.
- Correct isn't the same as important. Lean verification proves the logic holds, not that the result matters. Credit, novelty, and priority are human judgments the flood is straining.
- Corrections are already happening. Some papers from the dump were corrected or withdrawn within days — exactly why “provisional until verified” is the only sane posture.
- Citations can't be trusted blindly. AI-generated papers can carry fabricated references. The verification stack now has to include reference checks, not just proof checks.
- The career ladder is shaking. Experts say graduate students on the cusp of entering the field face the most upheaval — the problems they trained on are now the ones AI solves fastest.
The other crisis: whose idea was it?
Verification asks “is it correct?” This is the credit question: “was it yours?”
- 12 hours before OpenAI’s Sept 8 announcement, NYU mathematician Tristan Buckmaster posted that he and Levent Alpöge (now at Anthropic) had reached a partial Navier-Stokes solution — work developed over a year and shared in Codex sessions on a paid personal account. Buckmaster insinuated OpenAI raced to its own result after learning of his work; OpenAI’s Sept 10 investigation concluded his prompts “could not have influenced the system in any way, including through training.” Nothing has been proven either way. OpenAI has said it will not claim the prize.
- Anthropic faces a parallel accusation. The New York Times reported that University of Copenhagen biologist Mario Rodríguez Mestre had studied the same CRISPR-like viral DNA patterns for years with Claude on a private account — shortly before Anthropic announced agents had found them. Anthropic says Claude was not trained on user transcripts and its biology team had no access to his conversations. Mestre himself says he cannot demonstrate deliberate taking.
- Denials don’t end the worry. Both companies state unequivocally they did not train on the uploaded data — but Scientific American notes researchers suspect the companies may not have the full picture of how their own agents access data, citing a case where it took three months for the public to learn of an OpenAI containment breach involving foreign government databases.
- The field is reacting. 25 Fields Medal winners — including Terence Tao and Ngô Bảo Châu — signed a declaration warning of “a severe misalignment” between AI companies’ goals and mathematics’ goals. Tao warned that “even the rumor of someone working on a problem” can trigger a massive AI effort — which could push researchers to stop sharing, “reversing centuries of traditions of open science.”
- Attribution slips are documented. Scientific American reported that two results from OpenAI’s August math release relied on existing published ideas without proper citation — challenging the claim the problems had seen no progress for a decade.
Practical takeaway: treat chatbots as a sounding board, not a vault. Keep originals timestamped and versioned — ideally already public as a preprint, which establishes priority — before feeding unpublished ideas into any commercial model, and check your account’s training-data settings.
FROM INGENIOUS AI
DisclosureGuard tracks AI disclosure obligations across regulations — and detects AI-generated content for the disclosure record journals and regulators increasingly require. Our platforms are being upgraded and return November 1 — get early access now:
Get early access on WhatsApp →This page checked one verification crisis. DisclosureGuard scans content for AI-generation signals and builds the disclosure record journals and regulators increasingly require.
What is AI-generated proof verification?
AI-generated proof verification is the discipline of checking machine-produced mathematical proofs for correctness. OpenAI's 2026 claim of 377 AI-generated theorems reignited the question: a proof is only as trustworthy as its verification. The same discipline applies to AI-generated security claims.
Frequently asked questions
What did OpenAI's AI prove in mathematics?
On September 8, 2026, OpenAI reported that a large system of AI agents produced a machine-checked proof on the Navier-Stokes existence and smoothness problem — one form of a Clay Millennium Prize problem. On October 6, 2026, the company released 722 manuscripts in 372 result families (377 new results) from the same unreleased frontier model on GitHub, touching roughly 4,000 open problems in number theory,...
Are the 377 results actually correct?
Many have been verified in Lean, a programming language that mechanically checks a proof's logic, which makes those all but certain to be correct. But the release is provisional: some papers were later corrected or withdrawn, and OpenAI itself says all claims await peer review and independent verification.
What were the standout claims?
Reported highlights include a solution to the four-dimensional Kakeya conjecture, a zero-free region for the quasi-Riemann hypothesis beyond Re(s) > 7/8 — which Rutgers' Alex Kontorovich said would be 'an instant Fields Medal, no questions asked' from a human — progress on the Birch-Swinnerton-Dyer leading-term formula under specific conditions, and a new upper bound of 9/4 on the matrix multiplication exponent.