Anthropic’s Claude Mythos Finds Vulnerabilities in Cryptographic Algorithms

Anthropic researchers have used their advanced AI model, Claude Mythos Preview, to uncover mathematical vulnerabilities in two significant cryptographic algorithms: the HAWK digital signature scheme and a reduced-round version of the Advanced Encryption Standard (AES). These findings, though not immediately impacting current systems, highlight the growing role of AI in identifying potential weaknesses in foundational digital security frameworks.

HAWK, a candidate for post-quantum cryptography under consideration by the U.S. National Institute of Standards and Technology (NIST), was designed to withstand attacks from quantum computers. Remarkably, Claude Mythos was able to halve HAWK’s effective key strength in just 60 hours of analysis, a breakthrough that would typically take years of expert human review. This raises concerns about the robustness of cryptographic designs and the accelerated timeline AI tools bring to the cryptographic vetting process.

The second discovery involved a reduced-round version of AES, a widely used encryption standard. While full AES remains secure, Mythos developed an attack that improved the performance of a known cryptanalytic method by 200–800 times. This was achieved through an innovative optimization called the “Möbius Bridge,” enhancing the efficiency of existing attack techniques. Although this attack has no immediate application to the full AES algorithm, it underscores AI’s potential to refine cryptographic research at unprecedented speeds.

According to Anthropic, both breakthroughs were achieved with minimal human intervention, relying on Claude Mythos to autonomously explore, experiment, and validate its own findings. Each result cost approximately $100,000 in API usage fees, reflecting the resource-intensive nature of these experiments.

Broader Implications for Cybersecurity

The implications of these findings extend beyond academic cryptography. With models like Claude Mythos reportedly capable of breaking 65–86% of Tier-1 cryptographic schemes in benchmark testing, there are growing concerns about their dual-use potential. A recent report suggested that the U.S. National Security Agency (NSA) may be evaluating Mythos for offensive cyber operations, raising red flags about the ethical and geopolitical ramifications of AI-driven cryptanalysis.

Anthropic’s limited distribution of Mythos to vetted enterprise partners and critical infrastructure organizations reflects an understanding of the risks. The model’s pricing, starting at $10 per million input tokens and $50 per million output tokens, ensures that access remains restricted to entities with substantial resources and oversight capabilities.

Future Directions

Anthropic has announced plans to collaborate with academic institutions, such as ETH Zurich and Tel Aviv University, to build CryptanalysisBench—a benchmark framework for evaluating AI models in cryptography. This initiative aims to foster transparency and enable broader research into the cryptanalytic capabilities of frontier AI models.

However, the rapid advancement of AI in this field also raises questions about preparedness. What happens if AI models discover vulnerabilities in widely deployed cryptosystems? Anthropic has called for a multi-stakeholder conversation involving academia, industry, and governments to address these challenges proactively.

While the immediate impact of Claude Mythos’s discoveries is limited to research contexts, the pace and scale of AI-driven cryptanalysis signal a paradigm shift in cybersecurity. As AI models grow more capable, the cryptographic community will need to adapt quickly to safeguard the digital infrastructure that underpins global communications, finance, and commerce.

Image source: Shutterstock