AITech Trends

Anthropic’s Claude Finds Mathematical Flaws in Cryptography: Post-Quantum Scheme Halved, AES Attack 200-800x Faster

Anthropic has published new research showing that its AI model Claude Mythos Preview achieved two breakthroughs in cryptography. The first significantly weakens HAWK, a digital signature scheme built for a post-quantum world, effectively cutting its key strength in half. The second identifies a new way to attack round-reduced AES, the most widely used symmetric cipher, improving the speed of previous best attacks by 200 to 800 times. Anthropic stresses that neither result has a practical impact on today’s computer systems, but both show the potential for frontier AI models to help discover flaws in important cryptographic algorithms.

Key Takeaways

  • Claude Mythos Preview improved the best-known attack on the post-quantum signature scheme HAWK in 60 hours, halving its key strength
  • The research also found a new way to attack round-reduced AES, improving attack speed by 200-800x over previous best methods
  • Neither result affects production systems today; each result cost roughly $100,000 in API costs to develop

Two breakthroughs

The first result targets HAWK, a digital signature scheme that was built for a post-quantum world and is one of the third-round candidates under consideration in a 2022 NIST call for quantum-resistant systems. Despite surviving two rounds of expert human review over two years, Mythos improved the best-known attack on HAWK in just 60 hours, effectively cutting its key strength in half.

The second result concerns AES, the Advanced Encryption Standard adopted by NIST in 2001 and scrutinized more than almost any other encryption algorithm. Weaker variations of the algorithm are regularly studied to understand its robustness; Mythos found a way to break one such reduced version and eliminated one of the guesses an attacker needs to make, improving the speed of previous best attacks by 200 to 800 times.

No impact on production systems

Anthropic is clear that neither result affects today’s computer systems and no production software has to change. HAWK is only a candidate signature scheme and is not deployed, and the second attack targets a reduced version of AES rather than breaking the full cipher. Nevertheless, both results demonstrate the potential of frontier AI models to help discover flaws in cryptographic algorithms, both before and after real-world deployment.

Mostly autonomous, at a cost

Mythos Preview achieved these results mostly autonomously. Over the course of a week, one Anthropic researcher worked together with Claude to develop the HAWK attack, while another built a scaffold that allowed Claude to fully autonomously discover the AES attack. Each result cost roughly $100,000 in API costs. Anthropic also partnered with academics at ETH Zurich, Tel Aviv University and TU Berlin to build CryptanalysisBench, a benchmark for evaluating LLM cryptanalytic ability, and followed responsible disclosure procedures throughout.

Source: Anthropic Research