Isogeny Labs

Thesis

The move to post-quantum cryptography will replace the public-key cryptography in nearly every system in use. The new schemes are being deployed in new code against fixed deadlines, on cost estimates that have had little time to be tested, and that is why this lab exists.

Cryptographic hardness is estimated once and re-examined rarely. Someone computes the cost of the best known attack, publishes the figure, and the field cites it for decades. The algorithm is usually sound. The cost model ages, against hardware that does not stand still and against cryptanalysis that does not stop.

Most of the time nobody checks. Checking is engineering rather than mathematics, it produces no theorem, and the incentives in this field do not reward it. So the numbers that parameter choices, migration plans and procurement deadlines all rest on go unexamined for years at a time.

This lab builds the attacks and measures them. We implement known cryptanalytic methods against elliptic curve and isogeny based systems, report what they cost on hardware that exists now, and publish the code and the measurements so the figures can be checked by anyone who wants to. Where the estimate and the measurement disagree, the measurement is the finding.

Research

  1. Anthropic asked how the NIST curves were picked. Our AI agents found a 1-in-4,480 pattern.

    Autoresearcher · elliptic curve standards · NIST and ANSI X9.62 prime curves

    Anthropic's reconstruction of the NIST curve seeds left an open question: why were some curves that passed the safety checks published and others left out? Our agents found that in every batch, every published curve scores lower on one count than every curve left out, an ordering random picks would produce once in 4,480 tries. The count does not make any curve weaker, and the pattern could still be chance.

  2. Crowdsourcing an attack on ECC2K-130, to find out what it really costs

    Distributed attack · parallel Pollard rho · NVIDIA, Apple silicon and CPUs

    We are running a distributed attack on ECC2K-130, an elliptic curve challenge published in 1997 that nobody has solved, and we need machines. The experiment is not whether we break it but what the attack costs: Certicom priced it at 2.7 billion machine days in 1997 computers, and nobody appears to have checked that figure against hardware people actually own.

  3. A 29-year-old attack got 17 times cheaper on one NVIDIA Blackwell card, and the math didn't change

    GPU cryptanalysis · parallel Pollard rho · CUDA

    We made the program that attacks Certicom's ECC2K-130 challenge, open since 1997, run 17.2 times faster on one NVIDIA RTX PRO 6000 Blackwell without changing the mathematics. One card would now need about 4.3 years for the whole attack; 290 cards, about a week.

  4. We rented a few hundred machines and spent four weeks attacking a 29-year-old challenge

    Pollard rho · GPUs, CPUs and an FPGA · a distributed campaign

    We rebuilt the planned 2009 attack on Certicom's ECC2K-130 from scratch and ran it on rented GPUs, CPUs and an FPGA. After 486 million stored points, about 5.4% of the expected work, there is no collision, as the mathematics predicts. Finishing is a question of money, roughly $56,000 of GPU time at spot prices, not of mathematics.

  5. We spent eleven weeks pointing AI agents at cryptanalysis. Our tests kept passing code we knew was broken.

    Language-model agents · research harness · verification

    Eleven weeks of language-model agents doing cryptanalysis research, in a system built to catch confident wrong answers. There was no cryptanalytic advance. The finding is about verification: our own checker passed code we had deliberately broken, five times over, once because the code noticed it was being tested.

  6. Quantum computers and post-quantum cryptography

    Survey · Shor, Grover, the migration deadline, and isogenies

    What a quantum computer breaks and what it does not, why the move away from RSA and elliptic curves has to start before one exists, and where isogeny-based cryptography fits.

Solutions

Software and engagements for the move to post-quantum cryptography.

See products and services

Notes and reference

Notes
Eight pages on the mathematics our measurements rest on, from elliptic curves and isogenies to lattices, with worked examples computed by the code that builds this site.
Cryptanalysis reference
Twelve attacks that public-key parameters are sized against: what each needs, what it costs, and where it stops applying.
Schemes
A catalogue of the schemes in use and the ones replacing them, with twelve full pages, most of them with a worked example computed each time the site is built.