The number that moved
Between late May and July 26, 2026, a benchmark for how cheap a quantum computer would need to be to break Bitcoin and Ethereum's elliptic-curve signatures dropped from 10.75 billion to 1.496 billion — an 86% cut. That's not a press release exaggerating a lab result. It's the scoreboard from ECDSA.Fail, a competition run by Eigen Labs where more than 100 researchers, working with AI coding agents, iterated on circuit designs for attacking ECDSA, the signature scheme that secures most onchain private keys.
This is a story about what AI agents are actually good at right now: not "autonomous trading," but grinding through a well-defined, verifiable optimization problem faster than a lab of humans alone would. Worth separating from the unsourced "AI agents are coming for crypto" headlines, because this one has a scoreboard.
What the benchmark actually measures
The metric is logical qubits multiplied by Toffoli gates — a proxy for how large and how error-tolerant a quantum computer would need to be to run a working attack circuit against ECDSA. Lower is worse news for the defenders, in theory: a smaller number means less quantum hardware is needed to threaten the signature scheme protecting onchain funds.
The leading design that emerged from the competition uses 1,151 logical qubits with roughly 1.3 million Toffoli gates; later submissions pushed the gate count below 1 million. For context, that's in the same neighborhood as — roughly half, by one counting method — the benchmark Google Quantum AI published in March, though the two use different counting conventions, so treat that as directionally comparable, not an apples-to-apples drop.
The participant list is a small cross-section of the people who'd actually have to respond to this: Eigen Labs, Trail of Bits, StarkWare, Theta Labs, MultiVM Labs, and the Ethereum Foundation.
"Open Autoresearch" is the actual product here
Eigen Labs is calling the format "Open Autoresearch": humans and AI coding agents iterating against a shared, measurable target, with an automated verifier checking every submitted circuit design. No leaderboard-gaming, no self-reported numbers — a design either verifies or it doesn't.
That's the pattern worth stealing for anything you're building with agents: give the agent a target that a machine can grade, not a target only a human can eyeball. A chatbot that claims it improved something is a screenshot. A circuit that passes a verifier is a result. It's the same distinction SatoHub draws between self-reported and verified — the format matters as much as the claim.
The clock this feeds into
This benchmark isn't happening in a vacuum. The Ethereum Foundation has set a target of quantum-resistant transactions by December 2029. NIST's own guidance proposes deprecating classical public-key algorithms at the 112-bit security level after 2030, and disallowing them outright after 2035. A faster-moving attack benchmark shortens the runway between "theoretical" and "plan for it now," which is presumably why a serious research group ran the competition in public instead of sitting on it.
None of this means a quantum break of Bitcoin or Ethereum is imminent — the benchmark is a circuit-design cost estimate, not a working attack, and building the actual quantum hardware to run these circuits remains a separate, very hard problem. What it does mean is that the migration to post-quantum signature schemes has a research community actively pressure-testing the timeline, in public, with agents doing real work on the input side.
What to watch
Watch whether Eigen Labs runs another round of ECDSA.Fail and whether the benchmark keeps falling at this pace — a second 86% cut would be a different story than a one-time optimization win. Watch which post-quantum signature proposals the Ethereum Foundation actually commits to ahead of its 2029 deadline. And watch for more "Open Autoresearch"-style formats elsewhere in crypto security research — a verifier-in-the-loop competition is a much more interesting use of agents than another wallet-draining bot story.
Sources
- ▸[Decrypt — Morning Minute: AI Agents Cut BTC Quantum Attack Benchmark by 86%](https://decrypt.co/377948/morning-minute-ai-agents-cut-btc-quantum-attack-benchmark-by-86)