Artificial intelligence has crossed another threshold—and this time it wasn’t writing code or discovering software bugs. Anthropic says its experimental Claude Mythos Preview model has uncovered entirely new weaknesses in cryptographic algorithms themselves, a milestone that could reshape how researchers think about both AI and digital security.
The company’s Frontier Red Team revealed two significant cryptanalysis results that go beyond implementation flaws or coding mistakes. Instead of finding vulnerabilities in software, Claude helped identify mathematical weaknesses in the underlying algorithms that are designed to secure digital communications.
The findings do not put today’s encryption at risk. Bitcoin remains unaffected, modern versions of AES remain secure, and no production systems are vulnerable. Yet the implications are difficult to ignore. Just one year ago, frontier AI models were largely incapable of contributing meaningful original cryptanalysis. Today, they are producing research worthy of academic publication.
AI Moves Beyond Bug Hunting
Throughout 2026, AI has demonstrated remarkable capabilities in cybersecurity. Models have become increasingly effective at identifying software vulnerabilities, reviewing source code, and even generating working exploits under controlled conditions.
Finding implementation bugs, however, is fundamentally different from discovering weaknesses in cryptographic mathematics.
Modern cryptographic algorithms are among the most heavily scrutinized pieces of mathematics in existence. They undergo years of public analysis by academic researchers before they are ever considered for widespread adoption. Breaking—or even slightly weakening—such algorithms typically requires deep expertise in number theory, algebra, lattice mathematics, probability, and computer science.
According to Anthropic, Claude Mythos Preview has now demonstrated that frontier AI models can meaningfully contribute to this level of research.
The company’s Frontier Red Team described the work as attacks on the algorithms themselves rather than mistakes made by software developers implementing them.
That distinction is significant because implementation bugs can usually be fixed with software updates. Weaknesses in the mathematics behind an algorithm are much harder to address.
HAWK Loses Half Its Effective Security
The most notable result involves HAWK, a post-quantum digital signature algorithm that reached the third round of the U.S. National Institute of Standards and Technology’s post-quantum cryptography evaluation process.
Post-quantum cryptography aims to protect digital systems against future quantum computers, which could eventually break many of today’s public-key cryptographic methods.
HAWK was designed as one candidate capable of surviving that future.
Claude Mythos Preview reportedly discovered a previously unknown attack in approximately 60 hours that cuts HAWK’s effective security level roughly in half.
The result does not completely break the algorithm, nor does it affect deployed systems. HAWK has not become a widely adopted production standard.
Nevertheless, reducing an algorithm’s effective security by such a large margin represents a meaningful advance in cryptanalysis. For any candidate seeking standardization, newly discovered mathematical attacks significantly weaken its long-term prospects.
Anthropic disclosed the findings to the algorithm’s designers and relevant government partners before publishing the research.
Faster Attacks on Reduced AES
Claude also improved an existing attack against a seven-round version of AES-128.
At first glance, headlines claiming AI “broke AES” sound alarming. They are also misleading.
AES-128, the encryption algorithm protecting everything from banking systems to encrypted messaging applications and Wi-Fi networks, uses ten rounds of encryption.
The research targeted only a seven-round version—a deliberately weakened variant that cryptographers frequently analyze to understand an algorithm’s security margins.
Anthropic says Claude discovered improvements that accelerate the known attack by roughly 200 to 800 times.
Although impressive from a research perspective, the result does not threaten real-world AES encryption. Full AES-128 remains secure, and the newly discovered technique does not extend to the complete algorithm.
Instead, the work demonstrates that AI can contribute meaningful improvements to cryptanalytic research on problems experts have studied for decades.
Bitcoin Is Not Affected
The announcement naturally raises concerns within the cryptocurrency industry.
Fortunately, the immediate impact is essentially zero.
Anthropic explicitly stated that neither SHA-256 nor ECDSA—the two cryptographic foundations securing Bitcoin—are affected by the new discoveries.
SHA-256 continues to secure Bitcoin’s proof-of-work mining process, while ECDSA protects wallet signatures authorizing transactions.
Likewise, Ethereum and most other major cryptocurrencies are unaffected by the published research.
The HAWK attack targets an entirely different signature scheme that has never become part of mainstream blockchain infrastructure.
Similarly, the reduced-round AES research concerns symmetric encryption rather than the public-key cryptography used by cryptocurrency wallets.
For crypto investors, the findings should therefore be viewed as an indicator of future AI capability rather than an immediate security threat.
A New Era for Cryptanalysis
Perhaps the most important aspect of Anthropic’s announcement is not the specific algorithms involved but the speed at which AI produced the results.
According to the company, Claude required roughly 60 hours to identify the HAWK weakness.
The AES research took approximately one week.
Both projects required occasional human guidance rather than fully autonomous operation, but the overwhelming majority of the mathematical exploration was carried out by the model itself.
Anthropic estimates each research effort consumed roughly $100,000 worth of API computation.
Those costs remain substantial today.
Like virtually every major AI capability, however, computational expense has historically fallen rapidly as hardware improves and algorithms become more efficient.
If future generations become both stronger and cheaper, cryptographic research could accelerate dramatically.
AI Is Becoming a Research Partner
For decades, cryptanalysis has largely progressed through incremental advances produced by relatively small groups of academic specialists.
AI introduces an entirely different model.
Instead of replacing human cryptographers, frontier models may increasingly function as research collaborators capable of exploring enormous mathematical search spaces, testing hypotheses, generating proofs, and identifying unexpected attack paths.
Human researchers remain essential for validating discoveries, understanding theoretical implications, and determining whether proposed attacks have practical significance.
Claude did not independently revolutionize cryptography overnight.
But it demonstrated that AI can now contribute original ideas to one of the world’s most mathematically demanding disciplines.
That marks an important shift from earlier generations of language models, which excelled at summarizing existing knowledge but rarely produced novel scientific insights.
Security Researchers Gain a Powerful New Tool
Anthropic frames the research as a defensive capability rather than an offensive one.
Discovering weaknesses before malicious actors do has always been the foundation of modern cryptography.
The company says it privately disclosed its findings to algorithm designers, U.S. government agencies, and industry partners before releasing the results publicly.
This follows a growing trend among frontier AI developers to collaborate directly with cybersecurity organizations, software vendors, and infrastructure operators.
Rather than waiting for attackers to exploit vulnerabilities, companies increasingly hope AI can identify weaknesses early enough for researchers to strengthen systems before widespread deployment.
That philosophy aligns with Anthropic’s broader Project Glasswing initiative, which focuses on using advanced AI models to improve software security.
The cryptanalysis results represent an expansion of that effort from software vulnerabilities into the mathematics underlying digital security itself.
The Bigger Story Is AI’s Trajectory
Neither HAWK nor reduced-round AES represents an immediate crisis.
Production encryption remains secure.
Bitcoin remains secure.
Ethereum remains secure.
But the pace of progress is remarkable.
Only a year ago, frontier AI systems were not capable of producing original cryptanalytic results of this caliber. Today, they are contributing discoveries that would normally require experienced academic researchers working for weeks or months.
That trend matters far beyond the specific algorithms discussed this week.
Cryptography has always evolved alongside advances in mathematics and computing power. Artificial intelligence now appears poised to become another major force shaping that evolution.
Rather than replacing cryptographers, AI is becoming an increasingly capable research assistant—one that never tires, can evaluate vast numbers of mathematical possibilities, and continues improving with each new generation.
For the cybersecurity industry, that creates both an opportunity and a challenge. Defensive research can move faster than ever before, but so can the search for weaknesses.
The race is no longer simply between cryptographers and attackers.
It is increasingly becoming a race between AI systems working on both sides of the equation.