AI Is Breaking the Tools Crypto Needs Against Quantum
An artificial intelligence system developed by Anthropic has uncovered a mathematical weakness in HAWK, a post-quantum digital-signature candidate that had already survived two rounds of expert review.
Key Takeaways
- Each experiment cost roughly $100,000.
- Bitcoin’s existing signatures remain unaffected.
- Human verification took longer than discovery.
- Future cryptography must remain replaceable.
In a separate experiment, the same system produced an improved academic attack against a reduced version of AES, one of the most closely examined encryption standards in use today.
Neither result compromises Bitcoin, Ethereum, hardware wallets or production encryption. HAWK has yet to be deployed, and the AES work applies to seven rounds of a cipher that uses ten in practice.
What changes is how emerging cryptography may need to be tested. Using multi-agent systems, specialist mathematical software and $100,000 in API compute for each result, Anthropic compressed parts of the cryptanalytic process into days.
That development arrives as Bitcoin researchers consider a multi-year move away from signatures that could eventually become vulnerable to quantum computers. Any replacement will now face automated analysis alongside conventional human review.
Claude Reduced HAWK’s Security Margin
HAWK is a lattice-based scheme taking part in the third round of NIST’s additional post-quantum signature process.
Algorithms in that category are built around mathematical problems that known quantum methods are not expected to solve efficiently. Before Anthropic’s work, HAWK had undergone two years of public examination.
According to Anthropic’s research, Claude Mythos identified an overlooked symmetry in HAWK’s underlying lattice structure. Exploiting it produced a more efficient route to recovering the secret key.
New Anthropic research: Discovering cryptographic weaknesses with Claude.
Claude Mythos Preview has helped our researchers find weaknesses in cryptographic algorithms—the mathematical methods that are used to keep data private.
Read more: https://t.co/TYKLjb3Q7V
— Anthropic (@AnthropicAI) July 28, 2026
For HAWK-256, a deliberately small challenge parameter designed for cryptanalysis, the estimated attack cost fell from 264 operations to 238.
The larger proposed versions remain impractical to break, and the weakness stops short of every lattice-based design and the other candidates in NIST’s process.
HAWK would nevertheless need substantially stronger parameters to recover its intended security level. Anthropic estimates that its keys may need to double in size.
HAWK-512 already uses a 1,024-byte public key and a 555-byte signature. Increasing those figures would add storage, bandwidth and verification costs, eroding the efficiency advantage that made the scheme attractive.
On a blockchain, such overhead multiplies across transactions and is permanently replicated by nodes. A mathematically sound upgrade can still become impractical when its signatures consume too much block space.
The Attack Required More Than a Chatbot Prompt
Running semi-autonomously across a multi-agent research system, Claude took 60 hours to develop and test the HAWK attack.
The agents could search academic literature, write Python and Sage code, run mathematical experiments and share promising results. In one case, an approach initially rejected by one agent was continued by another until it produced the decisive breakthrough.
Human direction focused largely on organising the work, selecting tools and checking the resulting mathematics. Anthropic said the operator had a theoretical computer science background while lacking any specialism in lattice cryptography.
The experiment used $100,000 in API compute. That is a substantial expense, yet it sits within reach of well-funded research laboratories, security companies and sophisticated adversaries.
Work that once required assembling a team of scarce specialists can now be supplemented by rented systems capable of testing ideas continuously.
Bitcoin’s Quantum Proposal Still Needs a Signature Scheme
Bitcoin Improvement Proposal 361 outlines how the network could migrate away from the ECDSA and Schnorr signatures protecting transactions today.
Our earlier report of Bitcoin’s proposed defence against quantum attacks explains the planned sequence. Legacy coins would eventually be restricted to post-quantum destinations, followed later by tighter recovery rules for funds that had not moved.
The transition could span five years after activation, though BIP-361 remains a draft with no activation date, and it leaves the replacement scheme to a separate proposal.
Choosing that algorithm involves more than confirming resistance to known quantum attacks. Developers must weigh signature size, verification speed, hardware-wallet support, block-space demand and the possibility that later research forces stronger parameters.
HAWK illustrates the danger of treating a post-quantum label as a permanent guarantee. A design may remain unbreakable in practice while losing enough of its security margin to become inefficient or unsuitable for deployment.
Bitcoin’s eventual upgrade should therefore preserve room for another change. Building wallets and protocol rules around one supposedly permanent replacement could leave the network facing a second disruptive migration if that scheme is later weakened.
The AES Result Shows Broader Cryptanalytic Ability
For its second project, Anthropic turned to AES-128, which protects data through ten rounds of repeated mathematical transformations.
Claude worked against a seven-round research version under assumptions that require 2105 chosen plaintexts, conditions that make the attack unusable against real encrypted data.
The system still improved the previous best method by 200 to 800 times, depending on how the computational cost is measured.
Its main contribution was a fingerprinting technique called the “Möbius Bridge,” which removed a 256-way guess required by earlier approaches.
Reaching that result required an autonomous framework in which multiple agents proposed hypotheses, ran experiments and discarded failures. The complete run generated around one billion output tokens and consumed $100,000 in API usage.
AES has been studied for decades, so the result goes well beyond spotting an obvious programming mistake in unfamiliar software. It shows that AI systems can contribute original ideas to mature areas of mathematical cryptanalysis.
Verification Is Becoming the Slow Part
Claude produced the central AES insight within a week. Two human researchers then spent close to a month checking the mathematics and gaining confidence that the attack was valid.
Automated systems can pursue many speculative ideas at once, while qualified cryptographers must reproduce each credible result, determine whether it is genuinely new and assess its practical consequences. That gap between production and validation is where security teams will feel the strain.
A growing volume of machine-generated findings would include mistakes, rediscoveries of known work and a smaller number of serious vulnerabilities. Sorting them quickly may require as much investment as producing them.
Crypto projects will need clear submission channels, independent reproduction and strict severity standards. An unverified model output should not trigger a protocol change, though dismissing every machine-generated report risks overlooking a genuine flaw.
Crypto Can Use the Same Tools Defensively
Finding the HAWK weakness before deployment is a useful outcome. NIST can demand stronger parameters or reject the candidate before governments, wallets and financial systems depend on it.
Blockchain developers can apply similar systems to proposed signature schemes, zero-knowledge proofs, wallet recovery methods and cross-chain protocols before significant value sits behind them.
Older components deserve renewed scrutiny too. Many crypto projects rely on specialised designs that have received far less attention than AES or the leading NIST candidates.
Attackers can use comparable tools, and a decentralised network may need months to coordinate an upgrade once a weakness becomes public. Discovery speed and verification speed will both shape who holds the advantage.
What Bitcoin Developers Need to Prepare For
Bitcoin’s quantum roadmap now carries two uncertainties: when capable hardware may emerge, and which replacement scheme will continue to hold up under deeper analysis.
Several candidates should be tested in parallel, rather than building the entire migration around an early favourite. Evaluations must cover mathematical security, transaction weight, verification costs and compatibility with exchanges, custodians and hardware devices.
AI-assisted research can form part of that process alongside public competitions, academic review and implementation audits, with human experts confirming every finding before it influences consensus rules.
Software architecture matters as well. Replaceable cryptographic modules would give wallets and nodes a safer route to change algorithms if new research weakens the first standard adopted.
Then comes the coordination problem. A secure design must still be adopted by developers, exchanges, miners, node operators, hardware manufacturers and users holding coins that may not move for years.
AI Changes How Quantum Safety Must Be Built
Anthropic has not shown that AI can break Bitcoin’s existing signatures, and Claude performs none of the work of a quantum computer. Its impact lands earlier in the process, challenging proposed algorithms, exposing hidden mathematical structure and reducing security margins before those systems reach deployment.
Quantum computing remains the eventual threat to operational wallets, while AI reshapes the research environment in which Bitcoin’s defence must be designed and tested.
The HAWK result therefore says more about the standard Bitcoin’s future cryptography must meet than about any single candidate. A replacement must withstand quantum theory, conventional cryptanalysis and automated research systems capable of exploring ideas faster than human reviewers can follow.
- Disclaimer: Anthropic’s findings do not provide a practical attack against Bitcoin, Ethereum, full AES-128 or production cryptocurrency wallets. HAWK is not deployed, and its larger parameters remain impractical to attack using the reported method.
- Methodology: This analysis uses Anthropic’s official research, NIST’s HAWK materials and Coindoo’s examination of BIP-361. Statements about cryptographic flexibility and Bitcoin’s migration requirements are analytical conclusions drawn from those sources.
Kosta has reported on cryptocurrency markets and blockchain infrastructure since 2020, bringing over six years of hands-on experience in the crypto industry built through daily tracking of markets, trends, and emerging blockchain developments. Specializing in Bitcoin on-chain analysis, institutional ETF flows, and digital asset price action, his work at Coindoo has been cited by other news agencies and consistently covers market developments with a focus on data-driven reporting across Bitcoin, Ethereum, Solana, and XRP. Over the years, Kosta has contributed to multiple crypto media outlets in different regions, authoring over 6,000 articles across the sector. His reporting spans cryptocurrency markets and the broader fintech industry, tracking not only price action but also the technological and regulatory forces shaping the ecosystem. To support his analysis, Kosta actively leverages on-chain data and metrics from leading platforms such as Santiment, Glassnode, and CryptoQuant, enabling deeper, evidence-based market insights. He believes in the power of transparency and the data that underpins the blockchain ecosystem. His academic background in Marketing Management from Denmark further complements his analytical approach, adding a strong understanding of communication strategy and content positioning to his work.
