AI Agents Cut Bitcoin Quantum Attack Benchmark by 86%
The figure comes from the ECDSA. Fail preprint, an arXiv paper posted on September 9, 2026 with a frozen benchmark cutoff of July 26, 2026 at 09:21:55 UTC.
A crowdsourced competition that paired more than 100 human participants with AI coding agents reported an 86% reduction in a Bitcoin quantum attack benchmark, shrinking the score for a key secp256k1 circuit primitive from roughly 10.75 billion to about 1.496 billion, though the researchers stress the result is not an executable attack on Bitcoin.
The figure comes from the ECDSA.Fail preprint, an arXiv paper posted on September 9, 2026 with a frozen benchmark cutoff of July 26, 2026 at 09:21:55 UTC. It sits squarely at the AI-crypto intersection: an open leaderboard where AI agents iterated on quantum circuit designs relevant to breaking elliptic-curve cryptography. For related coverage, see Fed Experiment Shows How Bitcoin Rallies Attract New Crypto Buyers.
Launched by Eigen Labs in late May 2026 with contributors affiliated with Trail of Bits, StarkWare and the Ethereum Foundation, the competition drew over 400 promoted submissions across roughly eight weeks. The context matters as institutions from NIST to Galaxy weigh post-quantum risk, a theme running alongside broader market moves like Bitcoin’s slide below $77,000 in the current correction. For related coverage, see Bitcoin and Ethereum Just Saw Their Best Rally in Years as Spot, Futures, and ETFs Turned Bullish.
KEY POINTS
- The ECDSA.Fail preprint reports an 86.1% reduction in a benchmark score for a reversible secp256k1 circuit primitive relevant to Shor’s algorithm.
- The metric and the AI agents’ contribution are documented in the paper, but the study is observational and does not isolate AI’s causal role.
- The result does not establish a practical attack on Bitcoin; it omits error correction and full fault-tolerant resource estimates.
What the reported 86% benchmark reduction measures
The competition optimizes reversible secp256k1 mixed point addition, a primitive relevant to Shor’s elliptic-curve discrete-logarithm algorithm. Its score is defined as S = Q × T, multiplying peak logical qubit width by average executed Toffoli count, according to the preprint. For related coverage, see US Government Moves Alameda-Linked Bitcoin, Sparking Sell-Off Fears.
On that metric, the paper reports an 86.1% reduction, from approximately 10.75 billion to approximately 1.496 billion at the July 26 cutoff. The headline’s rounded 86% describes this benchmark score change, not a reduction in attack time, cost, or Bitcoin’s security margin.
Reported circuit benchmark reduction
86.1%
What changed in the benchmark
The baseline circuit used 2,715 logical qubits and 3,960,753 average executed Toffoli gates. The cutoff winner, commit 8e9c9a2, used 1,151 logical qubits and 1,299,453 average executed Toffoli gates, the two multiplicands behind the score drop.
Peak logical qubits: baseline → cutoff winner
2,715 → 1,151
The benchmark supplies one addend classically. A separate coherent, windowed-addition-compatible variant used 1,162 logical qubits and 1,684,161 average executed Toffoli gates, with an empirical success probability of 0.99809 across 100,000 random inputs.
A post-cutoff submission pushed the count further, to 1,321 logical qubits and 952,707 average executed Toffoli gates for a reported approximate score of 1.259 billion. Those figures belong to a later entry and should not be conflated with the cutoff winner’s 1,151-qubit design.
Where AI agents contributed
More than 100 participants working with AI agents produced the over 400 promoted submissions, framing this as an AI-agent-assisted optimization workflow rather than a purely human effort. The paper describes the trajectory as observational: no random assignment to human-only and agent-assisted groups, no common attempt budget, and no complete prompt or wall-clock log.
As a result, the study does not isolate AI’s causal contribution to the 86.1% improvement. Claims that AI independently drove the entire gain, circulated in some newsletter coverage, are unconfirmed and unsupported by the paper’s own methodology.
One further caveat sits in the scoring mechanics: 4.8% of scored source commits fall in a verifier-search bucket, where nonce search can lower the sampled score without a semantic circuit optimization. That is a share of commits, not a discount applied to the headline 86.1% figure.
What the benchmark could mean for Bitcoin security
The authors are explicit that the reported circuits do not constitute an executable attack or a complete fault-tolerant resource estimate. They omit physical error correction, architecture-dependent compilation, full non-Clifford accounting and validated integration into windowed Shor arithmetic.
This work improves a reversible point-addition primitive relevant to quantum attacks on elliptic-curve cryptography, but it does not constitute an executable attack or a complete fault-tolerant resource estimate.
— ECDSA.Fail paper authors (Jieyi Long et al.), as reported by Decrypt
What the result does and does not establish
A percentage reduction in a single circuit primitive cannot by itself determine attack feasibility or a timeline for Bitcoin exposure. The optimized metric covers one component of the arithmetic, and the logical qubit counts exclude the substantial physical overhead that real fault-tolerant hardware would require.
That distinction separates a benchmark improvement from a compromised key. No Bitcoin private key was cracked, no funds were affected, and the work does not establish an imminent Q-Day, contrary to how a headline figure alone might read.
Evidence needed to assess the risk
Assessing practical significance would require absolute resource estimates including error-correction overhead, explicit hardware assumptions, and validated integration into full Shor arithmetic, none of which the preprint claims to provide. Independent reproduction of the circuits and peer review remain outstanding, as this is a preprint that has not been externally replicated.
Institutional preparation, meanwhile, predates this paper. NIST’s IR 8547 is labeled an Initial Public Draft, published November 12, 2024 with a comment deadline of January 10, 2025, outlining a planned migration from quantum-vulnerable standards rather than any binding Bitcoin mandate.
On the industry side, Galaxy launched a Bitcoin Quantum Readiness Initiative in July 2026 with up to $5 million in developer grants, a research program and a Quantum Advisory Council. That is prior preparedness activity, not a demonstrated market reaction to ECDSA.Fail.
Bitcoin traded around $76,982 with a 24-hour decline of roughly 0.96% during the research window, and the broader Fear & Greed reading sat at 56, or Greed; neither reflects any measurable causal response to the paper. For the AI-crypto stack, the sharper question is whether agent-assisted circuit search meaningfully compresses the timeline for post-quantum migration, work that intersects with debates over onchain Bitcoin issuance and long-term protocol governance. Whether this reported improvement materially changes practical attack requirements remains unresolved pending absolute resource estimates and independent replication.
Disclaimer: This article is for informational purposes only and does not constitute financial or investment advice. Cryptocurrency and digital asset markets carry significant risk. Always do your own research before making decisions.
