Anthropic's Mythos AI found a critical flaw in HAWK, and the algorithm was pulled from the NIST standards race

Summary

Anthropic's Mythos AI model identified a previously unknown attack on the HAWK post-quantum signature algorithm during NIST's third round of evaluation. This attack effectively halved the algorithm's key strength, leading the developer to withdraw the candidate from the evaluations. The discovery required approximately 60 hours of computation and a cost of $100,000, combining existing mathematical methods rather than inventing new techniques.

Why it matters

This is one of the first high-profile cases where an AI system, operating semi-autonomously under human guidance, accelerated the cryptographic analysis of a standard candidate to the point of directly altering the NIST standardization process. It suggests that AI-assisted security research could reduce the time to find weaknesses in cryptographic building blocks from years to weeks. This situation could reshape how standardization bodies and enterprises assess algorithmic risk. The parallel discovery of an (impractical) attack on AES further underscores that AI is scanning not just experimental schemes, but the entire operational cryptography landscape.

Highlights

• HAWK's withdrawal means the remaining PQC signature candidates (ML-DSA, FN-DSA) will face less competition. This could narrow the standardization field.
• The cost of the attack (approximately $100,000, roughly 60 hours) is within the reach of well-resourced state or corporate adversaries. This lowers the barrier for targeted cryptographic analysis.
• Cryptographers, including a PQC researcher from Google, noted that there was already suspicion that HAWK possessed such security vulnerabilities. This limits the novelty of the AI's contribution.

Critical take

The article's framing of Mythos as a breakthrough is softened by its own reporting: the attack combines known techniques, the target was already suspected of being weak, and the AES finding is explicitly not described as practical in non-laboratory conditions. Calling this a 'critical security vulnerability' risks overstating what is, by the article's own account, an incremental advance that merely confirms existing suspicions faster and more cheaply than a human team could.