Key facts
- Anthropic's Claude Mythos Preview model discovered new attacks on cryptographic algorithms.
- One attack targets HAWK, a candidate for a U.S. federal post-quantum signature standard.
- The HAWK attack reduced the cost of recovering a secret key by approximately 67 million times.
- The model also accelerated an attack on a 7-round version of AES by 200 to 800 times.
- Anthropic researchers spent hundreds of hours verifying the AES findings.
- The company disclosed the results to relevant parties and coordinated the HAWK finding with NIST.
Anthropic announced that an unreleased version of its powerful AI model, Claude Mythos Preview, has discovered two previously unknown attacks on cryptographic algorithms. One of these attacks targets HAWK, a digital signature system designed to withstand future quantum computers and currently a candidate for a U.S. federal standard. The AI found a symmetry in HAWK's mathematics that reduced the cost of recovering a secret key from 2^64 operations to 2^38, a significant decrease that could necessitate doubling HAWK's key size, potentially diminishing its appeal as a post-quantum cryptography candidate.
Additionally, the model accelerated an attack on a 7-round research version of AES, a widely used cipher for securing data, by 200 to 800 times, surpassing a record set by cryptographers in 2013. While the AES finding required extensive verification by Anthropic researchers, the HAWK discovery was made in days. The company also noted that Claude Mythos Preview broke 13 rounds of LEA, a Korean national and ISO lightweight-encryption standard, in under an hour.
Anthropic has disclosed these findings to the algorithms' authors and relevant U.S. government and industry partners, coordinating the HAWK discovery with NIST. The company highlighted that human researchers are becoming a bottleneck in the cybersecurity community due to the rapid pace at which AI models can discover bugs.