A Paradigm Shift in AI Capabilities
For years, the intersection of artificial intelligence and cybersecurity has focused primarily on code implementation—finding the human-written bugs, buffer overflows, and configuration errors that allow hackers to breach software systems. However, a groundbreaking research announcement from Anthropic has revealed that frontier AI models have crossed a critical threshold. Utilizing its new experimental model, Claude Mythos Preview, researchers have demonstrated that artificial intelligence can autonomously discover fundamental mathematical flaws in the underlying cryptographic algorithms themselves.
According to the official research published by Anthropic, Claude Mythos Preview successfully engineered two major cryptographic breakthroughs. The first is a highly sophisticated attack that significantly weakens HAWK, a prominent digital signature scheme designed to secure data in a post-quantum world. The second is a novel mathematical attack on a round-reduced variant of the Advanced Encryption Standard (AES), the most widely utilized symmetric cipher globally. While Anthropic emphasizes that these findings do not threaten active production systems today, they represent a monumental leap in the capabilities of large language models (LLMs) to perform advanced, autonomous mathematical reasoning.
Dismantling HAWK: The Post-Quantum Vulnerability
As the horizon for cryptographically-relevant quantum computers shrinks, global standards bodies like the US Government’s National Institute of Standards and Technology (NIST) have been working for nearly a decade to standardize Post-Quantum Cryptography (PQC) schemes. These mathematical structures are designed to withstand attacks from quantum processors that could easily break classical encryption methods like RSA or ECDSA. HAWK, a lattice-based digital signature scheme, is currently a third-round candidate under close evaluation by NIST.
Despite HAWK having survived over two years of rigorous, expert-level human review, Claude Mythos Preview managed to improve the best-known attack on the system in just 60 hours of computational work. The AI model achieved this by identifying a previously unexploited mathematical symmetry—specifically, a nontrivial automorphism—within the lattice structure utilized by HAWK. While previous theoretical work suggested that finding such an automorphism could permit an attack, human mathematicians had not yet determined whether such a vulnerability was accessible within HAWK’s specific parameters.
By exploiting this symmetry, Mythos developed a faster exponential-time enumeration attack. The practical consequence is staggering: the attack cuts the effective key strength of HAWK in half. For instance, the expected computational cost to recover a key in the HAWK-256 configuration was previously believed to be 2^64 operations; Claude Mythos reduced this to just 2^38 operations. To restore the original security margins, developers would need to double HAWK’s key size, a modification that eliminates the very efficiency characteristics that made the scheme an attractive candidate for post-quantum deployment.
Cracking the AES Citadel: The Möbius Bridge
The second major breakthrough targets the Advanced Encryption Standard (AES), a symmetric block cipher adopted by NIST in 2001 and subjected to more intense scientific scrutiny than almost any other encryption algorithm in history. Cryptographers frequently study weakened, “round-reduced” variations of AES to estimate the safety margins of the full cipher. AES-128 utilizes 10 rounds of computational transformations; Anthropic’s researchers tasked Claude Mythos with attacking a 7-round variant.
Initially, the model resisted, generating outputs claiming that improving the cryptanalysis of AES-128 was mathematically impossible due to the sheer volume of existing literature. However, after researchers bypassed this safety-inhibition via targeted prompting, the model initiated a massive autonomous search. Over three days, working within a specialized execution scaffold, Claude Mythos generated approximately one billion output tokens, formulating and computationally testing hundreds of hypotheses.
The result was the creation of a highly sophisticated fingerprinting algorithm that the AI named the “Möbius Bridge.” This algorithm improves upon classical meet-in-the-middle attacks, which trade computational time for memory storage. Traditional attacks required an attacker to guess and enumerate 2^56 different intermediate values. The Möbius Bridge developed a mathematical invariant that bypasses this requirement entirely, reducing the necessary computational steps by a factor of 256. When combined with other optimization techniques discovered by the model, Claude’s attack proved to be 200 to 800 times faster than the previous state-of-the-art academic attacks on 7-round AES.
Autonomous Workflows and the Human Verification Bottleneck
The operational methodology behind these discoveries highlights both the immense power of agentic AI and the critical role of human oversight. The HAWK attack was developed semi-autonomously using a multi-agent sandbox environment similar to Claude Code, where multiple specialized AI agents collaborated, utilized computational tools like Python and Sage, and cross-referenced academic publications. Interestingly, the critical breakthrough was achieved when one AI agent prematurely dismissed an analytical path as infeasible, but a second collaborating agent salvaged the idea and fully exploited it.
Each of these cryptographic discoveries required substantial investment, costing approximately $100,000 in API usage fees. However, the research exposed a massive disparity in speed between AI generation and human verification. While Claude Mythos conceived and optimized the AES attack in a single week, it required two Anthropic researchers—who were not pre-existing experts in advanced cryptography—nearly a month of intense labor and hundreds of hours of study to verify that the mathematical proofs generated by the AI were indeed correct.
To broaden academic access and standardize the evaluation of these emerging AI capabilities, Anthropic partnered with leading scholars at ETH Zurich, Tel Aviv University, and the University of Haifa to launch “CryptanalysisBench.” This public benchmark packages various classical and modern ciphers, providing a standardized environment for researchers worldwide to test the cryptanalytic capabilities of frontier large language models.
Responsible Disclosure and Broader Horizon
In accordance with industry-standard security practices, Anthropic engaged in a rigorous responsible disclosure process. The findings regarding the HAWK vulnerability were shared directly with the algorithm’s original authors in June 2026, and a coordinated disclosure was subsequently delivered to the public NIST mailing list. Advanced copies of the research papers were also shared with US government agencies and key industry partners.
Beyond HAWK and AES, Anthropic revealed that Claude Mythos has already begun identifying vulnerabilities in other international standards. The model successfully executed a practical, end-to-end key recovery attack on a 13-round variant of the Lightweight Encryption Algorithm (LEA)—an ISO/IEC standardized cipher designed for low-power IoT devices—recovering the key in under an hour on a standard desktop computer. It also identified a key-recovery attack on a 6-round variant of the Serpent-128 cipher.
While none of these academic breakthroughs compromise the active, production-grade encryption protocols that secure global banking, defense, or consumer internet traffic today, they signal a profound shift. The design of future cryptographic standards will no longer be an exclusively human domain; instead, it will require a continuous, AI-driven arms race of automated stress-testing and synthesis to keep the digital world secure.