Researchers at Anthropic have discovered that artificial intelligence can identify vulnerabilities in cryptographic systems, raising questions about the future of digital security. The team’s Claude AI model independently developed new attack methods against two encryption targets: a simplified version of the Advanced Encryption Standard (AES-128) and HAWK, an experimental digital signature scheme under consideration by the U.S. National Institute of Standards and Technology for quantum-resistant applications.
While the findings generated significant headlines, security experts stress there is no immediate threat to consumers. The AI successfully attacked a deliberately weakened seven-round variant of AES-128 used by researchers for testing purposes, not the full ten-round version that protects real-world banking, shopping, and messaging systems. Against the simplified version, Claude’s technique proved 200 to 800 times faster than existing attacks, yet experts emphasize this breakthrough does not compromise the encryption safeguarding everyday digital communications.
The more significant achievement involves HAWK, where Claude identified a previously unexploited mathematical property that substantially weakens the algorithm’s security margin. Since HAWK remains under evaluation before widespread deployment, this discovery serves the intended purpose of the standardization process: identifying flaws before adoption rather than after. Cryptography specialists view this as evidence that AI could accelerate the evaluation of cryptographic systems across thousands of algorithms that human researchers lack time to thoroughly examine.
Rather than indicating AI has surpassed human cryptographers, the research demonstrates how artificial intelligence can apply expert-level analysis at scale, potentially strengthening future encryption standards before they become entrenched in critical infrastructure.
Originally reported by Livescience →
About the Author
