GPT-5.6 Sol Ultra credited in two unclonable-encryption preprints posted 3h18m apart
Separate teams submitted same-day arXiv papers claiming an efficient scheme without prior caveats, before peer review.
Two research teams used OpenAI’s newly released GPT-5.6 Sol Ultra to tackle the same open quantum-cryptography question and landed overlapping results within hours. Their arXiv submissions, credited to the model for core proof ideas, put “simultaneous discovery” and attribution norms under a brighter light.
Key Takeaways
- Two teams submitted separate arXiv preprints on the same unclonable-encryption result on the same day, with timestamps three hours and 18 minutes apart.
- Both manuscripts explicitly credit OpenAI’s newly released GPT-5.6 Sol Ultra for core proof ideas, with the researchers cleaning up, checking, and taking responsibility for the claims.
- The preprints claim an efficient unclonable-encryption scheme without the caveats or assumptions earlier approaches relied on, but neither paper had been peer-reviewed at the time described.
- Domain experts described the results as convincing while flagging new pressure points around credit, unequal access to tooling, and what gets left for graduate-student training.
A Same-Day arXiv Photo Finish for Unclonable Encryption
The collision was tight enough to have a timestamp. Prabhanjan Ananth (UC Santa Barbara) and Amit Sahai (UCLA) submitted their preprint to arXiv at 10:35 A.M. PDT. Seyoon Ragavan, an MIT Ph.D. student, submitted a separate preprint on the same result three hours and 18 minutes later.
The overlap surfaced before either manuscript was public. Ragavan sent his proof to Yao-Ting Lin, a UC Santa Barbara doctoral student, who was in a meeting with Ananth and heard about a second proof of the same result. Afterward, Lin emailed Ragavan, “We are definitely living in strange times.” Lin then connected the researchers once the duplication became clear.
The near-simultaneous discovery was not a pure coincidence. Ragavan and Sahai had both heard the same open question posed earlier in July 2026 at the Simons Institute for the Theory of Computing at UC Berkeley, then pursued it separately. The new ingredient was that both efforts pointed the same newly released model, GPT-5.6 Sol Ultra, at the same target.
What the Preprints Claim: Efficient Unclonable Encryption, Minus Prior Caveats
Unclonable encryption sits in a specialized corner of quantum cryptography, and the mechanism is the quantum no-cloning property. An unknown quantum state cannot be perfectly copied, which lets protocol designers aim for a stronger guarantee than classical encryption can offer in this specific setting: even if an adversary steals the ciphertext, they should not be able to split it into two usable “copies” that both decrypt once the key is later revealed.
The two preprints claim they can build an efficient unclonable-encryption scheme without the caveats or assumptions that earlier approaches needed. Prior work had established unclonable encryption was possible, but at the cost of inefficiency or additional assumptions for security. Anne Broadbent (University of Ottawa) and her then-student Sébastien Lord introduced a modern framework for unclonable encryption in 2019, which set the stage for the open question that resurfaced this month.
The catch for anyone trying to map this onto crypto-market narratives is process, not payload. Neither manuscript had been peer-reviewed at the time described, and the source material does not spell out the full formal statement of the open problem. That leaves the result in the “promising preprint” bucket rather than a confirmed security milestone with immediate implications for deployed systems.
Two Workflows, One Model: How GPT-5.6 Sol Ultra Was Used
Ragavan’s workflow was hands-on and iterative. He directed GPT-5.6 Sol Ultra in two-hour stretches, checked progress at each interval, and redirected when needed. OpenAI’s description of the Ultra system in the account was that it “coordinates four AI agents in parallel by default,” which Ragavan effectively treated as a way to keep multiple lines of attack running at once. After several rounds, the system produced a proof Ragavan believed was sound, plus a rough draft that he cleaned up and reorganized.
Ananth and Sahai used the same underlying model through a different interface and a different discipline. They worked off a bespoke UCLA system designed to help AI models pursue and critique candidate solutions. Their paper states the model produced the construction and the main proof ideas, while the researchers refined and verified the work and said they take responsibility for its claims.
That division of labor matters because it blurs novelty boundaries in a way academia is not set up to price cleanly. Ananth said the model’s first proposed construction resembled prior work, and he realized it close to posting time: “Just before we had to post online,” he said, “I remembered, ‘I’ve seen this scheme somewhere.’” Broadbent said she and collaborators had published essentially the same construction earlier in 2026. In that telling, the new contribution was not inventing the construction from scratch, but proving it meets the stronger security property researchers were seeking.
The researchers themselves framed this as the new norm for open problems. “Now the general mentality is: if someone mentions an open problem, the first thing is to see if GPT solves it,” Ananth said. “A lot of problems that we didn’t know how to solve are going to get solved in the very near future.” Ragavan put the speed in calendar terms: “This timeline thing is crazy,” he said. “It’s like two weeks and a day since this idea even formed.”
Signals for Crypto: Faster Cryptography Cycles, Verification Bottlenecks, and ‘Haves vs Have-Nots’
For crypto markets, the immediate signal is not that a quantum-cryptography primitive changed overnight. It is that research timelines can compress hard when a frontier model is explicitly credited for core proof ideas in two independent efforts that land within a few hours of each other.
That compression shifts the bottleneck to verification. If preprints arrive faster, the scarce resource becomes expert review, replication, and the slow work of checking security definitions and edge cases. Broadbent and quantum computing researcher Andrea Coladangelo said the new results appear convincing, but Broadbent also argued the literature itself may have made the target unusually “legible” to an AI system: “In hindsight, this was an obvious thing to look into,” she said. “There’s a body of literature, lots of conjectures, lots of schemes. You kind of feel like you gave it on a silver platter.”
The other pressure point is access and credit. Broadbent raised inequality concerns directly: “I have a lot of questions about haves and have-nots,” adding that automatable work is often what she would normally assign to graduate students. Ananth echoed the student angle: “I am happy that I am past being a student,” he said, “and I worry for the current crop.” Ragavan described the personal workflow break more than the career politics: “The way I do research now has nothing to do with how I did research two months ago,” he said. “The emotions are weird, but I think you’ve just got to adapt and roll with that.”
Near-term, the next concrete step is whether the two teams actually merge their overlapping work into a single version for conference submission, which they discussed after the overlap was identified. The more important step is whether peer review or public expert vetting confirms the claim of an efficient unclonable-encryption scheme without the prior caveats, or finds the kind of hidden assumption that tends to surface only when outsiders try to break the proof.
My Read: This Isn’t a Quantum Breakthrough Trade—It’s an AI Capability and Process Shock
The threshold that matters here is not whether unclonable encryption becomes a headline narrative for deployed crypto. It is whether “two teams, same open problem, same model, same-day preprints” becomes a repeatable pattern, because that would reprice how quickly theory can move from seminar question to public claim.
Because neither paper was peer-reviewed at the time described, this looks more like a sentiment catalyst about AI-accelerated cryptography than a confirmed security milestone. If the proofs hold up and the community converges on attribution and verification norms that can keep pace with multi-agent model workflows, the practical impact is a faster research cycle where validation, not ideation, becomes the binding constraint.
