In the race to secure digital infrastructure against quantum computers, the assumption has always been that some existing cryptographic algorithms will break. What was less certain is whether artificial intelligence could accelerate that breakage—and by how much. This week, a research model from Anthropic, Claude Mythos Preview, demonstrated exactly that: it uncovered a structural flaw in a proposed post-quantum signature scheme called HAWK, and did so in roughly 60 hours of compute time.
The HAWK digital-signature scheme is not yet deployed anywhere, and Bitcoin does not use it. But the timing could not be more relevant. Bitcoin developers and the wider cryptocurrency ecosystem are themselves scrambling to prepare for a quantum era, and the HAWK result is a stark reminder that the tools of cryptanalysis are improving faster than many migration timelines assume.
What is HAWK and why does it matter?
HAWK is one of several signature schemes designed to resist quantum attacks. Traditional digital signatures used across the internet and blockchain systems—such as ECDSA, RSA, and EdDSA—rely on mathematical problems that quantum computers could theoretically solve much faster than classical computers. The most famous of these is Shor's algorithm, which can efficiently factor large integers and compute discrete logarithms, threatening RSA and elliptic-curve cryptography.
Post-quantum schemes aim to replace those primitives with constructions based on lattice problems, hash-based signatures, or multivariate equations, none of which are known to be broken by quantum algorithms. HAWK falls into the lattice-based family, and it has been proposed as a contender for standardizing signatures in web and banking systems. Lattice-based cryptography is attractive because it offers strong security proofs and relatively efficient key sizes, but it is also mathematically intricate. The difficulty of solving lattice problems is the foundation upon which the whole scheme rests.
The AI attack did not break HAWK entirely. It found a flaw that effectively halves the security of its smallest key parameter set. In concrete terms, the work needed to crack that parameter set dropped from roughly 2^64 operations to 2^38 operations—a massive reduction. 2^64 was already considered borderline for classical attacks; 2^38 is trivial for a determined adversary with modern hardware. To put that in perspective, a modern consumer GPU can execute billions of operations per second, and a botnet or cloud cluster can easily scale to billions more. 2^38 is on the order of 275 billion operations, which a well-resourced attacker could complete in minutes or hours.
Even worse, the attack appears to make larger key sizes, which were introduced to compensate for weaker parameters, far less attractive. Designers of post-quantum schemes often rely on scaling key sizes to increase security. If an AI-assisted attack undermines that scaling, the entire security margin of a scheme must be called into question. Cryptographers typically aim for a security level of at least 2^128 operations, meaning an attacker would need 340 undecillion attempts to break the scheme. The HAWK finding shows that a scheme may claim such a level on paper, but in practice, its real-world security could be far lower.
The 60-hour attack
According to the details that emerged, the attack on HAWK was not a brute-force effort or a conventional mathematical reduction. It was a discovery process enabled by Anthropic's Claude Mythos Preview model. The AI system effectively searched for structural weaknesses, identified a vector of exploitation, and reduced the computational cost of a full break. The model was not specifically designed for cryptanalysis; rather, it used its general reasoning abilities to explore mathematical relationships within the scheme.
This is a departure from traditional cryptanalysis. In the past, breaking a cryptographic scheme required deep human insight, often built up over years of study. The most celebrated attacks—such as the breaking of MD5, or the development of the lattice-based LLL algorithm—are the products of human minds. AI offers a different path: it can generate hypotheses, test them rapidly, and iterate on millions of potential algebraic transformations. While human cryptographers can reason abstractly about structures, AI can act as a tireless assistant that explores a wider search space.
The cost of the computation was estimated at around $100,000. That includes cloud compute, model inference, and validation. For a national-state adversary or a well-funded research lab, $100,000 is negligible. The fact that an AI model can find cryptographic weaknesses in automated fashion, rather than a human cryptographer spending months or years, changes the threat landscape. It means that the cost of discovering certain classes of vulnerabilities is dropping, just as the cost of quantum computers is expected to drop over time.
This is not the first time AI has been applied to cryptography. Researchers have used neural networks to attack simplified ciphers, and there have been projects exploring AI-driven side-channel analysis. But the HAWK finding is notable because it targeted a serious, proposed signature scheme and produced a meaningful security reduction in a matter of days. It suggests that AI can accelerate not only the discovery of new algorithms, but also the discovery of new attack techniques.
What this means for Bitcoin
Bitcoin currently relies on ECDSA for transaction signatures and SHA-256 for proof-of-work and address hashing. Neither ECDSA nor SHA-256 will be directly affected by the HAWK result. The AI attack in question is specific to HAWK's algebraic structure, not to the mathematical foundation of elliptic curves. In other words, the vulnerability does not directly reveal a way to break ECDSA.
However, Bitcoin is not insulated from the broader issue. The long-term threat is real: a sufficiently powerful quantum computer could break ECDSA, allowing an attacker to forge signatures and steal funds. This is why the Bitcoin developer community has been debating the adoption of quantum-resistant signatures for years. The debate is not just about which post-quantum algorithm to choose, but about how to migrate an active network with trillions of dollars in value.
One of the biggest challenges is that Bitcoin's security model assumes that addresses are secure as long as the underlying hash function and signature algorithm are sound. But if a quantum computer were built today, it could, in principle, derive the private key from any public key that has been exposed on-chain. Bitcoin users who re-use addresses are particularly vulnerable. Over the years, many users have received funds at addresses that have been used multiple times, leaving the public keys visible on the blockchain. A quantum attack could target those exposed public keys.
Migrating Bitcoin to post-quantum signatures is not a simple upgrade. It would require a hard fork, and nodes and miners would have to coordinate on a new signature scheme. Many developers argue that the migration should have begun years ago, but there is no consensus on which algorithm to use. Some candidates, like SPHINCS+, have large signatures that would increase transaction sizes, possibly leading to higher fees and longer syncing times. Others, like Dilithium, offer smaller signatures but are less battle-tested in a blockchain environment.
The HAWK result shows that the timeline for quantum-resistant migration may be shorter than expected, not because quantum computers are arriving sooner, but because classical and AI-assisted attacks are improving rapidly. This is a subtle but crucial point: cryptography does not have to be broken by a quantum computer to be dangerous. It can be broken by better classical algorithms, and AI is accelerating the discovery of those algorithms. A scheme that was designed to be quantum-safe might be broken by an AI model running on today's classical hardware.
The broader post-quantum landscape
The HAWK finding is a setback for the post-quantum standardization ecosystem. The National Institute of Standards and Technology (NIST) has already selected several algorithms for post-quantum cryptography, including CRYSTALS-Kyber for key exchange and CRYSTALS-Dilithium for digital signatures. But the standardization process is ongoing, and alternatives like HAWK are still considered for specialized use cases. NIST has repeatedly emphasized that the security of these schemes is not absolute; it is based on current understanding, which can change.
The HAWK episode is a case study in why the cryptographic community needs to maintain a diverse portfolio of algorithms and be prepared to swap out primitives quickly. If a standardized algorithm is broken—whether by a quantum computer or by a novel classical attack—the consequences would be catastrophic for any organization that has adopted it. Diversity of algorithms provides a hedge against unforeseen weaknesses.
For blockchain networks, the challenge is even greater because of their decentralization. Upgrading consensus-critical cryptography requires coordination among thousands of independent actors. Even if a quantum-resistant scheme is standardized today, it could take years to implement and activate across the network. During that time, the scheme could be subject to further cryptanalytic efforts, including AI-assisted ones.
AI-assisted cryptanalysis adds a new layer of uncertainty. As language models and specialized machine-learning systems become more sophisticated, they may be able to analyze mathematical structures, search for patterns, and propose attack strategies far beyond what humans have explored. This doesn't just apply to digital signatures; it applies to hash functions, encryption schemes, zero-knowledge proofs, and other primitives. It also applies to the very assumptions on which blockchain security is built, such as the hardness of the discrete logarithm problem and the collision resistance of hash functions.
Some researchers argue that the entire cryptographic standards process will need to change. Instead of a slow, committee-driven approach, we may need continuous automated auditing of candidate algorithms. AI could be used not only to attack algorithms but also to defend them, by systematically searching for weaknesses before an adversary does. The HAWK attack is an example of how AI can be used as a tool for cryptanalysis, but the same tool could be turned around to validate the security of new designs.
The race is already underway
Bitcoin Core developers and contributors have already taken note of the HAWK result, even though it does not affect Bitcoin directly. The prevailing view is that any weakness in a proposed post-quantum scheme is informative, because it helps the community understand the security margins of different approaches. It also highlights the need for regular adversarial testing of the primitives that Bitcoin does use.
There are also more immediate concerns. The HAWK attack was relatively inexpensive and fast. If AI models can find cryptographic flaws in days, then existing schemes need to be tested with the same vigor. Bitcoin's own ECDSA implementation has been studied intensely for two decades, but that does not mean it is immune to a novel AI-generated attack. The Bitcoin community has a history of conservative decision-making, and the HAWK result may accelerate the adoption of stronger key derivation methods, or the use of address formats that require only a hash of a public key, not the full public key.
The Bitcoin improvement proposal pipeline includes discussions about quantum-resistant addresses, such as those based on hash-based signatures (Lamport signatures or SPHINCS+). But implementing these would be a major change to the network. There are trade-offs between security, usability, and decentralization. Some proposals would require users to hold new types of keys, while others would be transparent and backward-compatible.
The situation is analogous to the Y2K problem, but with cryptography: the risk is known, the fix is complicated, and the timing is uncertain. Just as the tech industry spent years preparing for a date change that, in the end, required a global effort, the crypto industry may need to prepare for a quantum event that has no fixed date. The difference is that Y2K had a hard deadline; quantum computing has a moving deadline that depends on technological progress.
An uncertain path forward
The HAWK episode has made one thing clear: the quantum-era transition is not just about waiting for quantum computers. It is about the ever-increasing power of classical computing, augmented by artificial intelligence. The same tools that can discover vulnerabilities can also help design stronger cryptographic systems. The challenge is to stay ahead in that arms race.
For Bitcoin users, the immediate takeaway is that their funds remain safe. Signatures based on ECDSA are still considered secure against classical attacks, and no practical quantum computer exists today that could threaten them. But the long-term picture is less certain. The HAWK flaw was found by an AI system in 60 hours. There is no guarantee that similar systems won't one day find flaws in the very foundations of blockchain security.
The race to quantum-resistant cryptography is no longer a theoretical exercise. It is being conducted in real time, with AI as a new participant. The outcome will determine not just the future of digital signatures, but the future of trust in decentralized networks. As the HAWK result demonstrates, the next breakthrough could come not from a lab, but from a language model crunching algebraic structures in a fraction of the time it would take a human. That is both a warning and an opportunity.
Source: Coindesk News