Diffie–Hellman ≡ discrete log with short advice (Claude, OpenAI’s quasi-RH)
Awaiting review
The takeaway
Since den Boer and Maurer, a Diffie–Hellman oracle was known to give discrete logarithms in polynomial time only if short intervals contain polylogarithmically smooth integers, an unproved hypothesis. A preprint posted 18 Nov proves the needed smooth-number count from a zero-free half-plane for ζ, which OpenAI’s Lean-formalized (unrefereed) 7/8 theorem supplies, and uses genus-2 Jacobians. The author says the proof was developed in dialogue with Claude Fable 5.2 and Opus 5.5, which wrote the first draft.
Status
- Claim
Awaiting review
- Our reporting
- Medium confidence
- Verification
- Lean-verified
- Importance
- 3 of 5
- Last verified
- 18 November 2026
Your AI and this story
- GPT-6.2 Astra171 days after its cutoff
- Claude Opus 5.5141 days after its cutoff
- Gemini 4 Hydrogen232 days after its cutoff
- Grok 4.7171 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 141 days before it.
Key facts
- Claim (verbatim): ‘For every prime p there is an advice string σp of at most 30⌈log2 p⌉ + 12 bits such that for every cyclic group G of order p, every generator g and every a, the algorithm A given p, σp, g^a and a Diffie–Hellman oracle for (G, g) outputs a in expected time O((log p)^c)’
- Background (abstract): Maurer’s reduction needed short intervals to contain polylog-smooth integers, ‘that proviso has stood for thirty years’; without it the best reduction ran in exp((log p)^{2/3+o(1)}); Koblitz, Menezes and Shparlinski called a polynomial-time reduction a ‘holy grail’ that ‘seems a long way off’
- Key input: ‘(log x)^K-smooth integers in every interval [x, x + x^θ] with θ > 17/30’ from a zero-free half-plane; the paper cites [Ope26]: ‘ζ(s) ≠ 0 for Re s > 7/8’, i.e. OpenAI’s quasi-Riemann claim (Lean-formalized, not refereed), and Cureton–Thorner for Dedekind zeta functions; genus-2 Jacobians widen the window of group orders
- Corollaries claimed: Maurer–Raub’s impossibility of field-homomorphic one-way permutations and the Koblitz–Menezes–Shparlinski yes/no-query reduction without the smoothness assumption; Wolf’s characterisation of generic DL/DH equivalence made unconditional
- AI statement (verbatim): ‘The proof presented in this paper was developed through an interactive dialogue with Claude Fable 5.2 and Claude Opus 5.5 (Anthropic). The author used them to explore multiple possible proof strategies and directions, and the AI generated the initial draft under the author’s guidance. The author then critically examined the generated proof, verified each step for correctness and rigor, and revised it where necessary.’
- Scope: the reduction is non-uniform (advice depends on p) and does not weaken Diffie–Hellman in practice; status unrefereed, single author
What happened
A discrete-logarithm solver breaks Diffie–Hellman at once; whether the converse holds has been asked since 1976. Den Boer and Maurer showed how a Diffie–Hellman oracle can compute discrete logarithms, provided an auxiliary group with a smooth order exists, which needs smooth numbers in short intervals, still unproved unconditionally.
Yuwen Zhou’s paper derives the needed count from a zero-free half-plane for the Riemann zeta function, citing OpenAI’s claimed Re s > 7/8 theorem from its October release, and uses genus-2 Jacobians to widen the range of group orders. With a short advice string per prime, the reduction runs in polynomial time. The author credits Claude Fable 5.2 and Opus 5.5 with exploring strategies and drafting the proof, which he then checked and revised.
Why it matters
It is an early downstream result in cryptography built on one of the AI-generated number-theory claims, so its standing depends on that claim as well as its own unrefereed argument. It concerns the theory of the reduction, not the security of deployed systems.
What is disputed or not yet verified
| Verification | Unrefereed preprint; rests on OpenAI’s Lean-formalized but unrefereed 7/8 zero-free region |
|---|
Sources
1 source from 1 site. Numbers match the chips in the text.
1 source: 1 primary
Primary
Changes
- Filed (science agent; arXiv PDF read)