Summary
- Anthropic announced that its unreleased Claude Mythos Preview model identified a new attack on HAWK, significantly reducing the effort needed to steal its smallest key from 2^64 to 2^38 operations.
- The model also accelerated an assault on a 7-round version of AES by up to 800 times, surpassing a record established by cryptographers in 2013.
- The cost of executing each of these results was approximately $100,000 in API usage, and Anthropic researchers dedicated several hundred hours to validate the AES findings.
Today, Anthropic revealed that a yet-to-be-released version of its advanced AI model, Claude Mythos Preview, has discovered two significant vulnerabilities in cryptographic algorithms, including one targeting HAWK, a contender for a U.S. federal digital signature standard.
HAWK is designed to secure digital signatures while ensuring protection against quantum computing threats. It was advanced to the third round of the National Institute of Standards and Technology (NIST) post-quantum signature competition in May, where it remains the sole lattice-based candidate.
Claude's discovery revealed a mathematical symmetry in HAWK that had previously gone unnoticed, enabling the recovery of a secret key at a reduced cost of 2^38 operations, making it roughly 67 million times less labor-intensive than before.
To rectify this vulnerability, HAWK would require its key sizes to be roughly doubled. Anthropic noted, "Unfortunately, doubling HAWK's key size eliminates many of the reasons making the scheme (as it currently stands) an attractive PQC signature candidate."
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
This adjustment is critical for blockchain technology, as signature size directly impacts block space and transaction fees. Thus, any blockchain seeking a quantum-resistant alternative must consider the bytes per signature. HAWK's appeal was its compact keys and quick signing capabilities, which would be compromised by the proposed fix.
For cryptocurrency holders, there is no immediate cause for concern; HAWK has not been implemented in any systems, and Bitcoin continues to operate on ECDSA, the pre-quantum digital signature scheme intended to be replaced by candidates like HAWK.
Before making its findings public, Anthropic shared both discoveries with the authors of the algorithms, as well as with U.S. government and industry stakeholders, while coordinating its HAWK findings with NIST.
Breakthrough in AES Cryptanalysis
The second vulnerability pertains to AES, the encryption standard that secures HTTPS traffic, encrypted drives, and backend systems of exchanges. The full AES-128 standard operates through 10 rounds of encryption, while Claude focused on a 7-round research version that has not seen improvements since 2013.
The testing conditions were intentionally challenging. Researchers prohibited the model from utilizing any of the five established families of AES cryptanalysis and instructed it to devise a sixth method, concluding the brief with a note that the initial differential attack did not surpass existing efforts—it created a new approach entirely. Claude also had access to notes from earlier attempts that had already explored about 200 unsuccessful attack variations.
Despite these limitations, Claude persisted. "On AES-128 r5/r6/r7 it found nothing because there's nothing easy to find; this is the most-studied block cipher in existence," the model relayed to researchers, as noted in transcripts released by Anthropic.
In the following three days, Anthropic received only three significant messages, including one stating: "no again the goal is that we have highly intelligent [sic] model as good top researcher, we want to find new attacks." Another message emphasized that Claude could not substitute AES for a simpler cipher.
Eventually, the model developed a technique referred to as a Möbius Bridge, which reduced the number of key bytes an attacker would need to guess from nine to eight. This refinement took additional days and a billion output tokens to finalize into a publishable version.
Another notable finding was Claude's ability to break 13 rounds of LEA, a national standard from Korea and an ISO lightweight encryption standard for mobile and IoT devices, in under an hour on a desktop, outperforming a previous best that required 2^98 plaintext pairs. The standard LEA implementation runs 24 rounds, meaning that currently deployed systems remain secure.
Longer Verification Process
The HAWK research paper candidly outlines the distribution of tasks, stating, "The majority of mathematical discoveries in this paper were AI-assisted. Human author contributions mainly consisted of directing, organizing and verifying AI work."
While Claude discovered the AES vulnerability in mere days, Anthropic's researchers invested several hundred hours to gain sufficient cryptographic knowledge to validate its effectiveness. This model had previously identified 271 vulnerabilities in Firefox during internal assessments.
"The cybersecurity community is now grappling with the fact that language models are able to discover so many bugs that the standard human processes (like vulnerability triage, verification, and remediation) struggle to keep up," Anthropic cautioned, indicating that human researchers could become a bottleneck in the process.
Additionally, Anthropic developed CryptanalysisBench, which includes 191 cipher-breaking tasks primarily sourced from NIST competitions. Models are required to submit a working attack script that either wins a formal security challenge or does not, without partial credit or human evaluation involved.
Mythos 5 successfully broke 85.7% of tasks with known solutions, compared to only 65.3% for the weakest model tested. However, against fully robust ciphers without published vulnerabilities, every model scored under 9%.