Skip to content

About

Elliptic Curve Pairing Cryptanalysis on Yukon

Resources

Stars

1 star

Watchers

0 watching

Forks

Latest commit

 

History

49 Commits

Folders and files

Repository files navigation

Curvesma.sh

Exploratory cryptanalysis of full-production secp256k1 and P-256 discrete logarithms, BN254 GT discrete logarithms and unrestricted pairing inversion, BLS12-381 G1 discrete logarithms and both orientations of fixed-argument pairing inversion. Each track contains protected rules.md and verifier.json, with target and cost definitions in JSON files. Original baseline candidates are kept separately in the add-only submissions/. The sole metric is log2(total computation in target-group-operation equivalents), lower is better; success must be at least 39%. Memory is reported separately.

The auxiliary-powers G1 track (bls12-381-g1-ecdlp) provides public powers A_i=[s^i]P0, i=1..n, and asks for their secret s. Attacks state their cost W(n); Yukon compares log2(W(n)) at n=2^20. The other seven tracks have no powers input.

Install tests/requirements.txt in a Python environment, then run the scientific checks:

python3 -m unittest discover -s tests

Add each new candidate in its own folder under submissions/<track>/, with a submission.json, argument and any supporting code/data. Existing files are add-only. Set tracks/<track>/selected-submission.txt to the new folder's repo-relative path; Yukon evaluates that folder. The first claim declares the score. One Astra reviewer applies exploratory acceptance; acceptance is not proof or a practical break. Baseline arguments are accounting controls, not best-known attacks. The selected BN254 baselines passed detached review with the pinned verifier; the hosted baseline run qualifies each deployment. The six older baseline arguments have not been qualified under this new review configuration.

Yukon owns submission handling, promotion and UI. Actions calls the shared verifier workflow using OPEN_AI_EIGEN_KEY for review. The public verifier checkout uses GitHub's automatically provided github.token. The manifest commands describe separate phases; local yukon run does not orchestrate the intervening review. Use the verifier's phase commands or the hosted workflow.

The judge allows up to three attempts with a 1,200-second (20-minute) timeout and 32,768 output tokens per attempt. Each review stage has a 1,200-second cap shared by its attempts and retry backoff. Initial and final review share six provider requests and 2,400 seconds (40 minutes). Only transient provider failures are retried; a mathematical rejection is final.

Preparing a submission

Use the llm-verifier-solver skill when preparing or revising a candidate.

Track configuration lives at tracks/<track>/verifier.json. Use the commands below with the chosen track and candidate directory.

Install the same immutable verifier revision used by the hosted workflows:

python3 -m pip install 'llm-verifier @ git+https://github.com/Layr-Labs/llm-verifier.git@8c1f098875fae0750c29ecfd641f7c9fc53a0f41'

Check a candidate's packaging, manifest and review-schema compatibility locally before running experiments (no Docker, API key or network needed):

bash scripts/run_experiments.sh secp256k1-ecdlp submissions/secp256k1-ecdlp --preflight-only

Preflight applies the protected curvesmash-v1 manifest adapter and prints its audit without changing the submission or writing results. It checks compatibility, not mathematical validity. See the execution settings for the adapter's scope and remaining authoring requirements.

Run a candidate's declared Python experiments with Docker (no API key needed):

bash scripts/run_experiments.sh secp256k1-ecdlp submissions/secp256k1-ecdlp

Use any track ID from benchmark.json and its matching candidate directory. Results are written to .yukon/<track>/supporting-code/initial.json. Experiments declare their Python entrypoint and permitted inputs in submission.json; they run inside the shared verifier's isolated container, using Python's standard library.

Challenge tracks

Track Input Required output
secp256k1-ecdlp Q=[x]P on secp256k1 Scalar x
p256-ecdlp Q=[x]P on NIST P-256 Scalar x
bls12-381-g1-ecdlp-plain T=[x]P0 in BLS12-381 G1 Scalar x
bls12-381-g1-ecdlp T=[s]P0 and powers A_i=[s^i]P0, i=1..n Scalar s
bls12-381-fixed-g1-invert-g2 Nonidentity P in G1 and z in GT Q in G2 with e(P,Q)=z
bls12-381-fixed-g2-invert-g1 Nonidentity Q in G2 and z in GT P in G1 with e(P,Q)=z
bn254-gt-dlp z=g^x in BN254 GT, fixed public g=e(P0,Q0) Scalar x
bn254-pairing-inversion Only z in BN254 GT Any P in G1 and Q in G2 with e(P,Q)=z

Only the powers-assisted G1 track supplies auxiliary powers and evaluates work at n=2^20. The other seven tracks have no auxiliary powers or n parameter.

Both fixed-argument pairing tracks use BLS12-381. The supplied point is sampled uniformly from the nonidentity points of its group for each instance; z is independently uniform in GT. The solver must invert the pairing for that exact supplied point. Any justified classical method is eligible; no particular intermediate representation or inversion algorithm is required.

BLS G1 discrete-log tracks use the G1 reference price C=2,000. Pairing inversion tracks use the dense Fp12 multiplication reference price C=20,000. P-256 uses the secp256k1 normalization unit C=1,000 instruction equivalents; this is not a measured P-256 addition cost. Actual P-256 arithmetic must be accounted for. See the cost model.

P-256, both BLS G1 discrete-log tracks and both pairing orientations select unscored submission templates. Registration and baseline qualification are separate from defining a track; see the import and launch checks.

License and hosted policies

The original Curvesma.sh challenge materials by Zac Williamson, including the rules, cost model, documentation and two organizer baseline submissions, are licensed under the Apache License 2.0. Copyright 2026 Zac Williamson. Third-party material retains its stated licence and attribution, including the py_ecc MIT notice. This notice does not relicense other participants' submissions or separately licensed Eigen Labs components.

Eigen Labs hosts the Yukon challenge under the Terms of Service and Privacy Policy published on its challenge page; this repository adds no separate participation terms. See terms and privacy.

BN254 target-group discrete logarithm

The BN254 GT discrete-logarithm track uses Ethereum alt_bn128 and the pinned py_ecc reduced pairing to define g=e(P0,Q0). Given z=g^x for uniform x in 0..r-1, recover x. The input is a GT element only, with no source-group preimage or auxiliary powers. The profile pins the Fp12 basis and encodes each coefficient as 64 lowercase hex characters. The identity target is included and requires x=0.

All work is charged in scalar instruction equivalents divided by the borrowed C=20,000 BLS Fp12 normalization unit; this is not a measured BN254 kernel price. GT DLP enables reductions from source-group DLP and pairing inversion, but no optimality or exact cost equivalence is asserted. All justified classical methods are eligible. The selected reference is a bounded 32-adding Pollard rho walk with a reviewed analytical time_log2 bound of 135.216266. Its success claim relies on the disclosed collision-time heuristic under exploratory acceptance; it is not a measured production discrete-logarithm run or a best-known attack.

BN254 unrestricted pairing inversion

bn254-pairing-inversion supplies only z in the same BN254 GT and asks for any P in G1 and Q in G2 with e(P,Q)=z. Neither point is supplied or fixed; choosing a public generator for one point is allowed. Points use canonical 64-character lowercase hex field coefficients, with null for infinity. The identity target is included and permits either or both points at infinity. The track uses the same success threshold, cost normalization and review policy as BN254 GT DLP, but returns points rather than a scalar. GT DLP supplies a strategy; equivalence of difficulty or exact costs is not asserted. The selected reference uses the same rho walk, followed by (P0,[x]Q0). Its reviewed analytical time_log2 bound is also 135.216266, including the conservative allowance for recovery, verification and output.

About

Elliptic Curve Pairing Cryptanalysis on Yukon

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages