Two research teams have independently solved a six-year-old problem in quantum cryptography using OpenAI’s language model GPT-5.6 Sol Ultra: efficient unclonable encryption. MIT doctoral student Seyoon Ragavan and cryptographers Prabhanjan Ananth and Amit Sahai submitted their proofs to arXiv on July 23, 2026, with a gap of about three hours between their submissions.
Two Teams Solve the Same Puzzle Independently
Unclonable encryption ensures that an intercepted ciphertext cannot be split into two separate, individually decryptable copies - a property that classical encryption cannot provide, but quantum states allow. An efficient, provably secure construction for this without additional security assumptions remained open for six years. The problem resurfaced in July 2026 during a talk at the Simons Institute at the University of California, Berkeley. Ragavan had attempted to solve it unsuccessfully years earlier and was surprised that it remained unsolved. He then set GPT-5.6 Sol Ultra to the task, letting the model work in stages of two hours each, checking the intermediate results, and intervening as needed before cleaning up and structuring the final construction. His paper “Efficient Unclonable Encryption from Pauli Eigenstates” describes a method where a random Pauli operator encodes a single classical bit across multiple qubits. Meanwhile, Ananth (UC Santa Barbara) and Sahai (UCLA) were working on the same question without knowledge of Ragavan’s attempt. They used a system developed at UCLA that alternates between generating solution proposals and critiquing its own intermediate steps. Their paper “Unconditional Unclonable Encryption” describes a construction without additional security assumptions with exponentially small distinguishing advantage for attackers. It was only after both papers had been submitted that UC Santa Barbara doctoral student Yao-Ting Lin noticed the overlap.
Scientific Community Discusses Utility and Fairness
According to Scientific American, the original AI idea initially aligned with an earlier work by Anne Broadbent’s group from 2026; the actual novelty lay in the subsequent proof of stronger security properties. Ananth sums up the new practice: when an open problem arises, one first checks if GPT can solve it. Ragavan commented on the timing with the words, “This timeline thing is crazy.” Physicist Anne Broadbent from the University of Ottawa views the approach more critically: automated proofs primarily address tasks that have so far been assigned to doctoral students - a question of equal opportunity in the training of the scientific workforce. The results of both teams are independently unverified, as peer review is still pending; both papers are currently being reviewed by the cryptography community and may be merged into a joint version. The debate thus fits into a broader discussion about how scientific journals and conferences will handle AI-assisted submissions in the future, particularly regarding authorship and review timelines.
It will be crucial to see if the pattern repeats: GPT-5.6 Sol Ultra had previously disproved a twenty-year-old statistical assumption and provided a purported proof for a fifty-year-old graph theory conjecture. It remains open how the cryptography community will handle simultaneously submitted, AI-assisted solutions to the same question in the future - particularly regarding the attribution of scientific authorship.
