OpenAI released hundreds of mathematical papers generated by an as-yet-unreleased internal model during the first part of October, placing them in the public domain. Spanning number theory, algebraic geometry, and related fields, the papers addressed questions that researchers had left unresolved for generations. Certain findings made contact with longstanding open problems in the field, while others brought new momentum to puzzles that had resisted resolution for extended periods.
A collection of around 719 manuscripts, distributed across 372 distinct families, pointed to a broader reconfiguration in how the field was taking shape. Lean proofs accompanied roughly four in ten of the highest-ranked outputs. The company itself flagged that certain informal assertions in the documentation might not be free of errors.
Only a few weeks before the release, OpenAI had put forward an assertion of having resolved certain aspects of the Navier-Stokes equations, a problem belonging to the Millennium Prize collection.
On 7 October, Scott Aaronson — a computer scientist — put forward the idea that AI laboratories run discreet internal tests of their newest models against cryptographic primitives, reasoning that uncovering any vulnerabilities through controlled channels would be preferable to those findings circulating without oversight. That suggestion drew a sequence of responses across a wider audience.
On the day that followed, Justin Drake — a researcher attached to the Ethereum Foundation — published a post on X calling on those across the blockchain industry to begin planning, in a measured way, for what he described as "bunker mode." The core of his concern centred on ECDSA — an elliptic curve cryptography standard governing how digital signatures are produced and validated — given that the protocol underpins transaction authentication across Bitcoin, Ethereum, and virtually every other widely adopted blockchain platform. Drake speculated — framing the view as his own opinion rather than established fact — that an AI system with advanced mathematical capabilities could eventually find a classical algorithmic route from public keys back to their private equivalents, and he put that window at months, not years.
Cryptography as a standalone subject produced no direct matches in the 719-item release, according to a review by Galaxy Research, while adjacent areas of study returned results the firm described as worth noting.
Ethereum co-founder Vitalik Buterin urged the sector to take the vulnerability on board without overreacting, cautioning that an immediate transfer of funds to fresh wallets fell outside what he could endorse. His recommendation centred on reducing reliance on encryption methods at risk from quantum computing, as well as those that may prove vulnerable to artificial intelligence. Hash-based signature schemes have found a foothold in portions of Ethereum's development plan.
Yehuda Lindell, head of cryptography at Coinbase, found no evidence that core elliptic curve assumptions had weakened, with years of accumulated research lending weight to conclusions about their durability. That internal assessment by Coinbase arrived at the same conclusion.
On October 9, Dan Boneh of Stanford joined partners from Galaxy Research at a discussion event the venture group organised and hosted. Addressing the audience, Boneh — who directs the applied cryptography research group at the university — described the findings as an advance in both mathematics and computer science, though one that, in his view, left no unresolved questions for the cryptography field. None of the hardness assumptions relied upon in practice were undermined by these findings, he said.
Justin Thaler, who holds a research partner role at the firm alongside a professorship at Georgetown, extended the argument further. The foundations of contemporary cryptography depend on a mathematical assumption — that certain computational problems have no fast solutions at all, not merely ones that researchers have yet to discover — and, as Thaler put it, the intelligence of any system, artificial or otherwise, cannot produce an algorithm where none exists. Thaler noted that the system has no way of locating an algorithm that does not exist. The P versus NP problem remains one of the most persistently unresolved questions in theoretical computer science. The output OpenAI delivered did not meet that standard in any of the areas measured, according to the assessments presented at the event.
Once a blockchain entry is recorded, that record cannot be altered or undone. Compromising ECDSA would leave every public key previously recorded on-chain without protection. Wallet holders whose transaction histories exposed public keys or showed repeated address use were flagged as facing the greatest degree of risk.
The approach Drake put forward centred on a phased transfer of holdings: the bulk of funds would shift to newly generated addresses whose public keys stay concealed behind a cryptographic hash, while a portion of the balance remains distributed across existing locations.
Matthew Green, a cryptography researcher at Johns Hopkins, drew attention to a wider conversation among security experts about whether advances in artificial intelligence could gradually erode the protections underpinning modern cryptographic standards, remarking that public-key encryption systems face a plausible threat of collapse. Recent mathematical work has led researchers to estimate that an encryption scheme rated at 128-bit protection may actually deliver closer to 100 bits of effective strength, a gap that would reduce the computational resources needed to mount an attack by a considerable margin.
Those specialists drew a clear line between this situation and the risks posed by quantum-based threats. Shor's algorithm gives quantum computers a way to tackle well-defined mathematical problems — among them integer factorisation and the discrete logarithm — that conventional machines handle far less efficiently. Some researchers argue that AI tools have the potential to hasten breakthroughs across the full breadth of mathematical fields.
At the time those conversations took place, researchers had yet to mount a working attack against RSA, ECDSA, or lattice-based cryptographic schemes in any live environment.
Defensive applications of the same technology are also within reach, according to researchers in the field. Tools for mathematically verifying smart contract code have found a practical home in automated theorem proving, which now plays an active role in that process. A collaboration between OpenAI and Paradigm took shape earlier this year around EVMbench, a testing framework designed to measure how well AI agents can locate, fix, and take advantage of weaknesses in software code.
Machine-learning tools have opened pathways toward novel constructions within the field of cryptography. Proof generation for unclonable encryption was accomplished through the application of computational models by one research team. Across two separate research groups, AI language model variants received attribution for foundational contributions in preprint papers uploaded to arXiv over the summer months.
Proprietary evaluations of internal models, run against established protocols behind closed doors, were described by people with knowledge of the matter who spoke to Aaronson. Those queries fall outside what publicly available models are set up to handle. Cryptography's absence from OpenAI's publicly released inventory raises the question of whether the omission was intentional rather than a reflection of any gap in the company's technical reach.
According to Haseeb Qureshi of Dragonfly, the threat to cryptographic security may stem not from quantum computing advances but from breakthroughs in conventional mathematics that could undermine foundational assumptions long treated as settled. Account abstraction and adaptable signature schemes — areas where Ethereum's development efforts remain active — are positioned to ease the path when protocol-level shifts become necessary, according to researchers working on the platform.
Researchers and security professionals have been directing growing interest toward post-quantum cryptography over a span of years. Ethereum's roadmap includes hash-based signature schemes, which researchers say offer resistance to quantum computing threats as well as a range of conventional cryptographic attacks. Lattice-based cryptography represents one additional avenue, though the HAWK algorithm was pulled from NIST's evaluation process after AI-assisted examination uncovered problems with the design.
One avenue Boneh raised involves putting advanced models to work attempting to break existing assumptions — and treating their repeated failure to do so as a form of evidence that those assumptions hold. One path forward, he suggested, would involve formalising those assumptions — codifying into an official structure the unverified premises that security arrangements have historically depended upon.
At the point those conversations took place, no cryptocurrency wallets had been emptied through AI-driven attacks, and no working proof had surfaced of any algorithm breaking ECDSA under real-world conditions. A number of projects in the cryptocurrency space have turned their attention to cryptographic agility — examining how readily their underlying primitives could be replaced if circumstances required it. Researchers and industry participants have indicated that developments over the coming months will provide further indication of whether this period represents a temporary fluctuation or the opening stage of a more lasting shift.
