News
Manuel B. Santos coauthors ECDSA.Fail
The new arXiv preprint studies how researchers and AI agents can improve quantum circuits through an open, evaluator-verified public benchmark.
ECDSA.Fail was released as an arXiv preprint on 9 September 2026. Manuel B. Santos, who leads cryptography research at PotomaQ, is a coauthor and contributed to major technical and empirical sections of the paper.
The work examines how people and AI agents improve reversible elliptic-curve point-addition circuits using a public evaluation process. These circuits are an important component of quantum cryptanalysis. The results concern evaluated circuit improvements and their resource costs, rather than a demonstration of complete key recovery on quantum hardware.
Manuel’s affiliation in the paper is MultiVM Labs. PotomaQ builds on work by that team and is developing AI research systems and quantum execution software.
What the paper studies
Shor’s algorithm for the elliptic-curve discrete logarithm problem spends most of its work adding points on the curve inside a reversible circuit. ECDSA.Fail treats the size of that circuit as an open optimization problem. The paper calls the format Open Autoresearch: humans and AI agents submit candidate circuits for the secp256k1 curve to a public leaderboard, an automated evaluator checks each candidate for correctness on sampled inputs, and accepted submissions are recorded with their provenance. The score is the peak logical qubit width multiplied by the average executed Toffoli count, so a submission has to account for memory and gate cost together.
The paper’s data cutoff is 26 July 2026, and version 1 was submitted to arXiv on 9 September 2026. Over the period it studies, the paper reports that participants reduced the benchmark score by 86.1 percent, with the best-scoring circuit at the cutoff using 1,151 logical qubits and 1,299,453 average executed Toffoli gates. It also documents the improvement chronology and the research practices observed among participants, and reports that AI agents complemented human judgment.
These are logical-circuit results under the benchmark’s accounting conventions; compiling such a circuit to a fault-tolerant architecture is a separate step. We read the paper as evidence that a well-specified, machine-checkable objective lets a mixed community of researchers and AI systems make measurable progress on a hard component of quantum cryptanalysis.