GPT-5.6 сгенерировал доказательство, решившее 6-летнюю задачу криптографии

Cryptographers Prabhanjan Ananth of the University of California, Santa Barbara, and Amit Sahai of UCLA posted a paper titled "Unconditional Unclonable Encryption" to arXiv on July 23, 2026, closing a problem that had stood before quantum cryptography for six years. The encryption construction and the key proof ideas were generated entirely by the Codex agent running on the GPT-5.6 Sol Ultra model.

What exactly did GPT-5.6 do

GPT-5.6, via the Codex agent, produced not just a fragment of the derivation but the entire encryption construction and the main ideas behind the proof of its security — Ananth and Sahai state this directly in a separate section at the end of the introduction. This paragraph is being discussed no less than the mathematical result itself: the language model independently proposed a research-level construction rather than performing a mechanical calculation.

• Publication date — July 23, 2026, preprint on arXiv (2607.21551)
• Authors — Prabhanjan Ananth (UC Santa Barbara) and Amit Sahai (UCLA)
• Model — GPT-5.6 Sol Ultra, operating through the Codex agent
• The problem had remained open for quantum cryptographers for about six years
• Paper topic — "Unconditional Unclonable Encryption"

What unclonable encryption is

Unclonable encryption is a scheme in which an encrypted message physically cannot be copied: the property relies on the quantum no-cloning theorem. The word "unconditional" in the title means that security is proven without computational assumptions — not under the assumption that some problem is hard to break, but unconditionally.

The unconditional variant was precisely the sticking point. According to the paper, a full-fledged construction of such encryption had remained an open problem for about six years — researchers, as Habr puts it, had "one bullet in the chamber," meaning a single working approach, and finding a second one had long proven elusive.

Why this matters

The paper is a notable case where a large language model produced not a numerical answer but a research-level mathematical construction and proof idea for an open problem. This shifts AI's role from "assistant for routine work" to co-author of substantive scientific work.

"The construction and the main proof ideas were generated entirely by the

Codex agent on the GPT-5.6 Sol Ultra model," states the section at the end of the introduction of Ananth and Sahai's paper.

According to arXiv, the preprint was posted on July 23, 2026, and is undergoing the platform's usual open community review — peer review in specialized journals still lies ahead. But the very fact that the model proposed a working construction for a problem that specialized groups had been struggling with for six years has already made the section about GPT-5.6 the center of discussion.

What this means

The boundary between "AI computes" and "AI proves" is shifting: GPT-5.6 proposed a construction and proof ideas for a problem where the human community had lacked a second approach for six years. If the result holds up after peer review, it will be one of the first high-profile examples of a language model serving as the author of a substantive part of a scientific paper, rather than a verification tool.

Frequently Asked Questions

Which AI model generated the proof?

The encryption construction and the main proof ideas were generated by the Codex agent on the GPT-5.6 Sol Ultra model — the paper's own authors state this in a section at the end of the introduction.

Who are the authors of the paper and where was it published?

The authors are Prabhanjan Ananth (University of California, Santa Barbara) and Amit Sahai (UCLA). The preprint "Unconditional Unclonable Encryption" was published on arXiv on July 23, 2026 (2607.21551).