AI Produces Two Proofs for the Same Cryptography Problem
GPT-5.6 Sol Ultra generated solutions for two separate teams simultaneously, raising questions about scientific discovery processes.
1Solutions submitted 3 hours apart
MIT's Seyoon Ragavan and UC researchers Prabhanjan Ananth with Amit Sahai independently solved the same quantum cryptography problem using OpenAI's GPT-5.6 Sol Ultra. Their arXiv submissions were only 3 hours and 18 minutes apart. This incident has reignited debates about AI's role in scientific discovery and the nature of independent research in the age of powerful language models.
2Understanding unclonable encryption
The problem centered on 'unclonable encryption', which leverages fundamental properties of quantum information theory. Unlike classical data, unknown quantum states cannot be perfectly copied. The framework developed by Anne Broadbent and Sébastien Lord in 2019 ensures that even if a message is intercepted, it cannot be split into two usable copies that both reveal the original content when decrypted.
3Divergent approaches, convergent results
Ragavan guided GPT-5.6 through two-hour working sessions, checking progress at each interval. The UC team used a custom system designed to help AI models critique potential solutions. Both methods relied on the model generating core proof ideas, which the researchers then refined and verified. This demonstrates multiple viable pathways for AI-assisted mathematical discovery.
4Priority disputes in AI-era science
Ananth recognized that part of their solution had appeared in prior work. The new papers proved this construction offered stronger security guarantees. Broadbent noted the result was 'obvious in hindsight.' The case highlights growing challenges in determining originality and attribution in AI-augmented research.
5The spark at Simons Institute
Both teams first heard about the open problem during a talk at UC Berkeley's Simons Institute for the Theory of Computing. This shared starting point created a natural experiment showing how AI is changing collaboration dynamics in theoretical computer science. The incident provides concrete data on how quickly modern tools can advance solutions to longstanding problems.
6Turning point for academic processes
Ananth observes, 'Now when people hear an open problem, their first instinct is to ask GPT.' Researchers predict many previously intractable problems will soon yield to AI-assisted approaches. This shift may fundamentally alter traditional academic practices from peer review to paper writing, requiring new norms for AI's role in scholarship.
7Solving decades-old problems
Models like GPT-5.6 have gained attention by solving 80-year-old mathematical conjectures. Theoretical computer scientists are systematically testing these tools' proof-generation capabilities. The quantum cryptography success demonstrates AI's potential for abstract mathematical reasoning, particularly in specialized subfields where human expertise is scarce.
8New era of scientific discovery
The near-simultaneous solutions illustrate the accelerated pace of AI-assisted research. Ragavan expressed astonishment at the timeline: 'It's been just 15 days since this idea formed.' Scientific priority systems and collaboration models will need to adapt to this new velocity of discovery, where traditional concepts of independent work may no longer apply.