Two Teams Asked the Same AI. Their Proofs Landed 3 Hours Apart.

TL;DR

On July 23, 2026, two research groups uploaded preprints to arXiv resolving the same open problem in quantum cryptography: an efficient, unconditionally secure unclonable encryption scheme. The submissions landed three hours and eighteen minutes apart. Both credit OpenAI's GPT-5.6 Sol Ultra with finding the core proof ideas, both were driven by researchers who heard the problem posed at the same Berkeley workshop earlier that month, and neither team knew the other was working on it. One of the papers says so in its arXiv comment field, in writing, as a matter of record.

What actually got proved

Unclonable encryption is a genuinely quantum idea with no classical analogue. A classical ciphertext is a string, and strings copy for free, so nothing stops an adversary from handing identical copies to two accomplices and waiting for the key to leak. Quantum states do not copy. The goal is a ciphertext that is a quantum state such that an adversary who holds it cannot split it into two pieces that both decrypt once the key shows up later.

Picture a concert ticket printed on a soap bubble. You can pass it to exactly one person, and any attempt to photocopy it for a second person pops it. That is the intuition; the technical content is proving that the pop is unavoidable rather than merely inconvenient, and proving it without leaning on any unproven computational hardness assumption.

That last clause is the open problem. Prior constructions bought unclonability with assumptions or with inefficiency. Both new papers claim the unconditional, efficient version:

• Efficient Unclonable Encryption from Pauli Eigenstates, by Seyoon Ragavan of MIT, gives what its abstract calls the first plain-model, one-time, information-theoretically secure, efficient unclonable encryption scheme for one classical bit. The construction picks a uniformly random phase-free Pauli on n qubits and encrypts the bit as a random eigenstate, bounding the probability that both receivers recover the bit at 1/2 + O(2^(-n/2)).
• Unconditional Unclonable Encryption, by Prabhanjan Ananth of UC Santa Barbara and Amit Sahai of UCLA, gives an unconditional construction of an information-theoretically secure one-time private-key unclonable encryption scheme for one-bit messages, with efficient encryption and decryption and exponentially small unclonable-indistinguishability advantage.

Read those two sentences again. They are describing the same milestone, written independently, on the same afternoon. Both also appear on the IACR Cryptology ePrint Archive as entries 2026/1509 and 2026/1511.

Same model, two very different harnesses

The convergence was not quite spontaneous. Both Ragavan and Sahai heard the question raised in July at the Simons Institute for the Theory of Computing at UC Berkeley, so the starting gun fired for everyone in the room at once. What differed was the machinery each side pointed at it.

Ragavan worked the model directly, running it in roughly two-hour stretches, checking where it had gotten to, redirecting it, then cleaning up the resulting proof himself. Ananth and Sahai did not converse with the model at all in the usual sense. They ran Codex on top of the same GPT-5.6 Sol Ultra through a bespoke UCLA system built to make the model generate candidate solutions and then attack its own candidates, with the humans refining and verifying whatever survived.

So this is not one workflow producing one answer twice. It is two structurally different research harnesses, one conversational and one adversarial, landing on the same result within a single working day.

The acknowledgment is the actual news

Papers have quietly used AI assistance for a while now. What makes this pair different is that the credit is on the permanent record, in the metadata, unhedged. Ragavan's arXiv comment reads:

GPT-5.6 Sol Ultra found this proof in an extended conversation with the author and drafted a preliminary version of this paper. The author is fully accountable for the correctness of this paper.

Two sentences, and they cleanly separate the two things people usually conflate. The model found it. The human owns it. That is a workable norm, and it is more than most of the discourse has managed in two years of arguing about it.

The Ananth and Sahai paper carries a comparable note crediting Codex running GPT-5.6 Sol Ultra for steps of the argument, driven by a harness the authors built. Ananth's summary of the resulting culture shift, to Scientific American, is blunt:

Now the general mentality is: if someone mentions an open problem, the first thing is to see if GPT solves it.

Ragavan puts the personal version more starkly: the way he does research now, he says, has nothing to do with how he did research two months ago.

Why a three-hour gap matters more than the proof

Simultaneous independent discovery is one of the oldest patterns in science. Newton and Leibniz on calculus, Darwin and Wallace on natural selection, and roughly every priority dispute since. The pattern is not new. The interval is.

Historically that interval ran months to decades, which is exactly why the machinery of academic credit works at all. Journals, submission stamps, and conference deadlines all assume a gap wide enough for a human process to resolve who got there first. Three hours and eighteen minutes is not a gap. It is a rounding error, and the bottleneck in theory research stopped being ideas and briefly became the arXiv upload queue.

If you build with these models, the practical read is this. A frontier model plus a competent domain expert can now close a freshly posed open problem in a research field over a few sessions, which means any problem publicly stated at a workshop is effectively being attacked in parallel by everyone who heard it, with near-identical tooling. Being first is now much closer to a coin flip than it used to be, and the thing that actually differentiated the two teams here was the harness, not the model.

The caveats, straight

• Both papers are preprints. Neither has completed peer review, and unconditional security claims in quantum cryptography have a history of surviving scrutiny only after several people have hammered on them.
• The two constructions are not identical. The titles alone signal different framings, and the papers may end up regarded as complementary rather than redundant.
• The convergence had a common cause. Both principals heard the problem at the same Simons Institute event in July, so this is parallel work on a shared prompt, not two people spontaneously typing the same question into a box.
• Human verification did real work in both cases. The model produced core ideas and constructions; the researchers refined, checked, and signed for them.
• This is one data point, and it is a self-reported one. Nobody has independently replicated what the model contributed versus what the humans contributed.

Key Takeaways

• Two groups posted proofs of the same open problem in unclonable encryption to arXiv on July 23, 2026, three hours and eighteen minutes apart, at 10:35 and 13:53 PDT.
• Both credit OpenAI's GPT-5.6 Sol Ultra with the core proof ideas, and Ragavan's arXiv comment states it explicitly while keeping full accountability with the human author.
• The two teams used structurally different workflows: direct conversation in two-hour stretches versus a custom UCLA harness running Codex to propose and critique candidates.
• The result itself is real work: an efficient, plain-model, information-theoretically secure unclonable encryption scheme for one bit, with cloning advantage falling off exponentially in the number of qubits.
• The credit system assumes a discovery gap measured in months. A three-hour gap is the part of this story that does not have an existing answer.
• Both papers remain unreviewed preprints, and the shared Simons Institute workshop is a plausible common cause for the timing.

Sources: arXiv:2607.21811, arXiv:2607.21551, IACR ePrint 2026/1509, IACR ePrint 2026/1511, Scientific American, Simons Institute for the Theory of Computing