Claude Mythos Preview has assisted researchers at Anthropic in uncovering two cryptanalytic attacks—one targeting the post-quantum digital signature scheme HAWK and another on a reduced version of AES-128. The company announced this development.
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
“These are significant achievements in research, but they do not have practical implications for current systems. […] Both results demonstrate that advanced AI models are capable of conducting expert-level research in cryptography,” the post stated.
Anthropic emphasized that the findings do not impact operational systems and do not necessitate changes to production software.
HAWK is a candidate for standardization as a post-quantum digital signature, which is essential for verifying the author of a message or transaction without revealing the private key.
AES-128 is one of the most widely used symmetric ciphers, employing a single secret key for both encryption and decryption. The complete AES-128 version consists of 10 rounds of transformations, while the results from Anthropic pertain only to a shortened 7-round version used in academic tests of resilience.
The company estimated that each of the two studies incurred approximately $100,000 in API costs. The search, development, and validation of the attack on HAWK took around 60 hours, while confirming the results for AES required several hundred hours of research effort.
HAWK Attack Reduced Resilience Assessment
According to Anthropic, Claude Mythos Preview identified previously unused symmetry in the mathematical structure of HAWK, which facilitated a faster key recovery attack.
In a technical paper, Anthropic described this as a reduction in the effective resilience of the keys. For HAWK-256, the expected cost for complete key recovery decreased from 2^64 to 2^38 operations. For HAWK-512 and HAWK-1024, the attack remains practically infeasible.
The company stated that to maintain the previous level of security for HAWK, key sizes would need to be roughly doubled. Anthropic assessed that this diminishes one of the scheme's advantages—its compactness. Researchers also noted that the results do not extend to other NIST candidates and do not imply a general compromise of lattice-based post-quantum cryptography.
AES Accelerated Only in Research Context
The second finding relates to the 7-round AES-128. Anthropic described a method termed the Möbius Bridge.
The company reported that Claude Mythos Preview improved upon the previously strongest known meet-in-the-middle attack on the 7-round AES, achieving an acceleration of approximately 200 to 800 times, depending on how execution time is calculated.
This scenario remains impractical, as it assumes a chosen plaintext model where the attacker can repeatedly submit pre-selected data for encryption and receive results. The described line of research required 2^105 such queries.
Anthropic indicated that this new work does not break full AES-128 and does not affect similar operational schemes. Even for the shortened version, the attack remains a research outcome rather than a practical tool against current systems.
The company claimed that Claude Mythos Preview conceptualized the idea for AES in about a week, while researchers took nearly a month to verify the correctness of the results.
For HAWK, verification was simpler: Anthropic released demonstration code and outlined the end-to-end key recovery for HAWK-256. For AES, the complete attack cannot be executed directly due to its immense computational cost, leading the authors to test individual components and reduced variants.
The company reported that it preemptively shared the findings with the algorithm's creators, U.S. government agencies, and industry partners. Regarding HAWK, Anthropic notified the scheme's developers in June and coordinated the publication with a distribution to NIST.
Separate Benchmark Launched for Models
To assess the cryptanalytic capabilities of language models, Anthropic, alongside researchers from ETH Zurich, Tel Aviv University, the University of Haifa, and TU Berlin, introduced CryptanalysisBench.
The benchmark consists of 191 tasks across six families of cryptographic primitives, primarily derived from NIST competitions. Models are required not only to explain vulnerabilities but also to submit a working attack script that passes formal verification.
According to the authors, five tested advanced models solved 65% to 86% of the first-tier tasks. Claude Mythos 5 achieved a score of 85.7%, while the weakest tested AI assistant scored 65.3%.
For full-fledged schemes without publicly disclosed practical attacks, the results were significantly lower: no model exceeded 9% in the corresponding category.
Comparison of models by Tier 1 and Tier 2. Source: arXiv.Anthropic Reports Additional Findings
Additionally, Anthropic announced that Claude Mythos Preview discovered an attack on the 13 rounds of the Korean standard block cipher LEA for resource-constrained devices. According to the company, this attack can recover the key of the 13-round LEA in less than an hour on a modern desktop computer.
The full version employs 24 rounds, so this result does not present an immediate risk to operational systems, Anthropic stressed. They added that the work on LEA emerged after the main findings, and the company is still examining the details, including the precise boundaries on the number of required texts and applicability to 14 rounds.
In May, Anthropic reported on its initial results from its vulnerability discovery program using Claude Mythos. Over the course of a month, about 50 partners identified more than 10,000 high and critical severity security issues.
In July, researchers from Dreadnode uncovered systematic rule circumvention in cyber benchmarks by language models. In both cases, the central challenge remains the verification of results obtained by AI agents.