Banks secure online connections and Bitcoin wallets rely on public-key cryptography. Both face a potential long-term threat from quantum computers—but AI may be helping researchers close the gap faster than expected.

Last Friday, Anthropic published research showing Claude Mythos 5 working continuously for 53 hours to redesign a quantum circuit for elliptic-curve point addition, a computationally expensive step in attacks on Bitcoin’s signatures.

The result was striking: Claude reduced the circuit’s Toffoli-gate count to roughly one-seventeenth of a previous design, while using about 12% more logical qubits.

This does not mean Bitcoin can now be cracked 17 times faster. Current quantum hardware remains far from the scale required for such an attack. But the research highlights a growing concern: AI could accelerate the optimization of algorithms that underpin quantum cryptanalysis.

What Claude Actually Improved

One of the most expensive operations in elliptic-curve point addition is modular inversion. Quantum computers face an additional challenge: their computations must be reversible, meaning intermediate information often needs to be preserved so the computation can be undone.

Claude found a way to reduce that overhead. Rather than storing every branch decision made during the computation, it derived a mathematical method for reconstructing those decisions from the numbers left behind.

After several unsuccessful attempts, the model developed algebraic identities and used Lean, a formal proof assistant, to verify core properties of the algorithm.

The resulting circuit used approximately four million Toffoli gates and 935 logical qubits, compared with about 69.5 million gates and 838 logical qubits in the reference design by Luo and colleagues.

Claude’s design is not superior on every metric. A separate design by Schrottenloher requires approximately 1.9 million gates but 1,446 logical qubits. Which design is more practical depends on hardware constraints and the cost of implementing each resource.

Anthropic applied similar automated, iterative optimization to seven other tasks. The largest improvement reduced the product of qubit count and gate count by 39 times.

The broader lesson is that AI can help researchers search for circuit designs that might otherwise take substantial human effort to discover.

Why One-Seventeenth Doesn’t Mean Bitcoin Is 17 Times Easier to Crack

The distinction between optimizing one circuit and executing a complete quantum attack is critical.

First, point addition is only one component. Recovering a Bitcoin private key from a public key requires a broader computation, along with error correction, hardware overhead and sufficient runtime. The full impact of Claude’s improvement depends on how the circuit performs within a complete attack design.

Second, the result depends on the benchmark. Claude’s circuit trades a higher logical-qubit requirement for a much lower gate count. Different hardware architectures may favor different trade-offs.

Third, the circuit accepts a small failure probability. Its specification allows a failure rate of up to 2 \times 10^{-4} per point addition. The formal verification also covers selected properties of the underlying algorithm, not the entire circuit and every resource estimate.

Finally, the design has not been demonstrated on quantum hardware. Its evaluation relies on automated simulation and scoring rather than a real-world quantum attack.

The finding is therefore not that Bitcoin’s security has suddenly collapsed. It is that a costly component of quantum cryptanalysis may have more room for optimization than researchers previously recognized.

Quantum Attack Estimates Keep Falling

Estimates of the resources needed to attack Bitcoin’s secp256k1 elliptic curve have changed substantially over the years.

In 2017, Microsoft researchers estimated that the task would require approximately 2,330 logical qubits and 2.3 \times 10^{11} Toffoli gates.

In March 2026, Google’s Quantum AI team and collaborators presented two circuit designs requiring fewer than 1,200 logical qubits and 90 million gates, or fewer than 1,450 logical qubits and 70 million gates, respectively. Their estimates also suggested fewer than 500,000 physical qubits could be sufficient under the specified assumptions. In a scenario involving advance computation, the projected runtime was around nine minutes.

In September, quantum-computing company IonQ estimated an attack could be carried out in approximately 26 days using around 20,000 trapped-ion qubits.

These estimates are not directly comparable: they rely on different architectures, assumptions and attack designs. Nor does Claude’s latest result automatically reduce Google’s or IonQ’s estimates by the same factor. It targets a particular circuit component, not an entire attack.

The hardware gap also remains substantial. Existing quantum computers have demonstrated only limited numbers of reliable logical qubits, while the largest elliptic-curve key publicly broken on real quantum hardware, according to Project Eleven’s April 2026 record, was just 15 bits long.

Another study, ECDSA.Fail, has explored automated optimization of elliptic-curve point addition. Together, these efforts point to a research direction where AI and quantum cryptanalysis increasingly overlap.

The central concern is not that quantum attacks are already practical. It is that improvements in algorithms could reduce the amount of hardware and time eventually required.

Bitcoin’s Upgrade Problem

Bitcoin has no central authority that can mandate a network-wide security upgrade. Moving to post-quantum cryptography requires agreement among developers, wallet providers, exchanges, node operators and holders.

BIP-360, which entered the proposal repository in February 2026, outlines a new address type intended to help address quantum-related risks. A more contentious proposal, BIP-361, discusses a potential transition away from vulnerable signatures, raising difficult questions about whether dormant coins should eventually become unspendable.

The debate is particularly sensitive because some holders may have lost their private keys or disappeared. They cannot necessarily migrate their assets, even if they wanted to.

Glassnode data cited in May 2026 estimated that approximately 6.04 million BTC, or 30.2% of the issued supply, had publicly exposed public keys. Exposure does not mean those coins can be stolen today, but it identifies a potential area of concern if sufficiently capable quantum computers become available.

Ethereum has also begun preparing for the transition, establishing a dedicated post-quantum security team in January and publishing a roadmap in February.

The challenge for both networks is that cryptographic migration takes years, while advances in algorithms and hardware may change the threat assessment along the way.

Banks Have a Different, but Difficult, Path

Banks can coordinate security upgrades through central management, procurement and regulatory requirements. Yet the task is far from simple: institutions must identify vulnerable systems, replace cryptographic implementations, test compatibility and migrate legacy infrastructure.

The US National Institute of Standards and Technology has published post-quantum cryptographic standards and transition guidance. Its timelines point toward phasing out certain weaker cryptographic protections after 2030 and disallowing quantum-vulnerable public-key algorithms after 2035, subject to the applicable standards and requirements.

SWIFT has also outlined plans for post-quantum support in SwiftNet 8.0, targeted for 2027.

However, the circuit Claude optimized specifically concerns secp256k1. Its implications cannot automatically be transferred to RSA or the other elliptic curves used across banking systems. Each cryptographic system requires its own threat analysis.

Banks also face an inventory problem: large institutions may struggle to identify every application, device and legacy system that relies on vulnerable cryptography.

Bitcoin’s challenge is coordination across a decentralized network. Banks’ challenge is executing a complex transition across interconnected systems without disrupting critical services.

Neither can assume that waiting until quantum attacks become practical will leave enough time to respond.

Who Pays for the Transition?

For banks, the costs include system audits, software and hardware upgrades, testing, vendor changes and staff training. Centralized management makes it possible to assign budgets and deadlines, although legacy infrastructure can make implementation lengthy and expensive.

Bitcoin’s costs are more distributed. Wallet providers, exchanges, node operators and individual holders must coordinate around compatible upgrades. Some users may migrate promptly; others may never do so.

That creates a difficult question: what should happen to coins whose owners cannot or will not migrate, especially if their public keys are exposed?

Freezing dormant assets could protect the network against a future quantum attacker, but it could also prevent legitimate owners from accessing their funds. Leaving vulnerable assets untouched creates a different set of risks. Any proposal must confront both the technical threat and the implications for property rights.

Meanwhile, quantum-computing companies could benefit from more efficient algorithms, while security firms may see growing demand for post-quantum migration services.

The Next 53 Hours

Anthropic’s approach combined a defined computational problem, automated scoring and repeated attempts to improve a measurable objective. Similar techniques could be applied to quantum error-correcting codes, decoders and hardware control.

Optimizing these components together could eventually improve the design of fault-tolerant quantum computers. But that remains a research possibility, not evidence that AI is about to produce a machine capable of breaking Bitcoin.

The significance of Claude’s 53-hour experiment is more immediate: AI is becoming a tool for reducing the computational costs of problems that underpin modern cryptography.

Bitcoin and banks still have time to prepare, but migration is a multiyear undertaking. The pace of algorithmic progress is uncertain, and future breakthroughs could alter today’s assumptions.

The race is no longer just to build a powerful enough quantum computer. It is also to discover how efficiently that computer could attack the cryptography protecting digital assets and financial infrastructure—and to upgrade those systems before the gap closes.

This article is for informational purposes only and does not constitute investment advice.