Anthropic AI Finds Cryptographic Flaw, Raising Questions Over Quantum-Safe Security

Anthropic has revealed that its Claude Mythos Preview AI model uncovered a major weakness in HAWK, a proposed post-quantum digital signature algorithm designed to secure future online banking, web payments, and other digital transactions. The discovery highlights how advanced AI could accelerate cryptographic research and reshape the race toward quantum-resistant security.

According to Anthropic, Claude spent roughly 60 hours analyzing HAWK at a computing cost of about $100,000. The AI identified an attack that reduces the effective key strength of HAWK's smallest parameter set from an estimated 2^64 operations to 2^38 operations, despite the algorithm having passed two years of expert review and multiple rounds of human evaluation.

Digital signature algorithms verify the authenticity of messages and transactions, making them essential for financial systems, websites, and blockchain networks. While Bitcoin and Ethereum rely on elliptic curve cryptography, Anthropic emphasized that its findings do not affect either cryptocurrency because HAWK is not part of their current or proposed security framework.

Bitcoin's proposed quantum-resistance roadmap, outlined in BIP-360, instead relies on three quantum-safe algorithms already standardized by the U.S. National Institute of Standards and Technology (NIST), providing multiple fallback options if vulnerabilities emerge. Meanwhile, BIP-361 argues that improving cryptographic attacks are shortening the timeline for migrating to stronger security, a trend Anthropic's research appears to reinforce.

The company also disclosed a second breakthrough, improving attacks on a deliberately weakened version of the AES encryption algorithm by 200 to 800 times. However, its AI achieved less than a tenfold improvement against Poseidon, the hash function widely used in zero-knowledge proof systems powering blockchain rollups and privacy-focused protocols.

Anthropic noted that Claude initially dismissed the AES challenge as too difficult before succeeding after several additional prompts, generating nearly one billion output tokens over three days. Human researchers then spent almost a month verifying the results.

The announcement comes as privacy-focused cryptocurrency Zcash recently activated a new shielded pool designed to remain secure in a future with quantum computers, underscoring how AI-driven cryptographic breakthroughs could influence the next generation of blockchain security.