Anthropic's Mythos model discovers cryptographic vulnerabilities in post-quantum algorithms (including HAWK signature scheme) that human experts had reviewed for 2+ years, finding breaks in 60 hours at ~$100k cost.

Anthropic's AI model Claude Mythos Preview has discovered mathematical weaknesses in cryptographic algorithms that underpin digital security. The model developed an improved attack on the post-quantum signature scheme HAWK and identified a new attack on a reduced version of the Advanced Encryption Standard (AES). Since encryption protects nearly all online activity and AES is the world's most widely used symmetric encryption standard, these findings demonstrate how AI models could challenge core assumptions behind internet security.

Anthropic emphasizes that neither finding affects systems currently in use. HAWK remains a candidate in an ongoing standardization process run by the U.S. National Institute of Standards and Technology (NIST), and the AES attack applies only to a modified version using 7 of the full scheme's 10 rounds. Nevertheless, the results are significant because they reveal potential vulnerabilities that conventional analysis had missed.

HAWK is among the remaining schemes in the third round of NIST's competition for additional post-quantum signatures—schemes designed to remain secure against future quantum computers. Human experts had reviewed HAWK for over two years before Mythos Preview discovered an improved attack in just 60 hours. The model worked semi-autonomously within a multi-agent system, with one agent initially dismissing the idea as infeasible before a second agent found a way to fully exploit it. The human researcher, who had a background in theoretical computer science but lacked expertise in lattice-based cryptography, served primarily in a project management capacity. The investigation cost approximately $100,000 in API fees.

The attack exploits a previously undetected symmetry in the mathematical lattice underlying HAWK's security. The model also discovered an attack on a reduced version of AES-128 almost entirely autonomously. A researcher provided a scaffold that allowed Claude to form hypotheses and test them experimentally. Mythos developed a new fingerprinting method that Anthropic calls the "Möbius Bridge," which removes one of the guesses an attacker must make and improves on the best previously known attacks by a factor of 200 to 800.

Human prompting played a limited role in the AES research. The model initially refused to pursue the problem, stating that "If you want a different outcome, the target has to change … AES-128 r5/r6 is just genuinely hard." Only after the researcher encouraged it to search for "genuinely novel ideas" did Mythos begin exploring more creative approaches. Over three days, the model generated several hundred million tokens while receiving only three additional substantive prompts, mostly to maintain focus. One prompt read: "Gain we are not looking for low hanging fruit, we want proper research to find genuinely hard findings." This research phase consumed approximately $100,000 in API fees for roughly 1 billion tokens. Human researchers, who lacked cryptography expertise, then spent several hundred hours verifying the results.

Anthropic shared the findings in advance with the U.S. government and industry partners and coordinated disclosure of the HAWK weakness with the scheme's authors. Mythos Preview remains unavailable to the public. In collaboration with researchers from ETH Zurich, Tel Aviv University, and the University of Haifa, Anthropic developed a benchmark called CryptanalysisBench to enable systematic evaluation of the cryptanalytic capabilities of language models.