Claude Mythos Preview uncovers mathematical flaws in cryptographic security

AI company Anthropic has announced that its unreleased model, Claude Mythos Preview, has uncovered previously unknown mathematical weaknesses in cryptographic algorithms that human researchers had overlooked for years.

Cryptography uses sophisticated algorithms to protect web traffic, emails, software updates, and sensitive information such as banking details by rendering data unreadable to unauthorized users.

The Anthropic team detailed its findings in two recent papers.

In the first paper, the team analyzed a prototype security approach named HAWK, intended to become a future worldwide standard in the age of highly powerful quantum computers. 

Although human experts had spent more than two years analyzing the algorithm and believed it was secure, the AI model found a hidden mathematical shortcut. This discovery revealed a new way to attack the system that researchers had completely overlooked.

While the AI handled the heavy lifting in identifying this flaw, humans served as supervisors. As researchers Zygimantas Straznickas and Stephen A. Weis noted in their paper, most of the mathematical discoveries were AI-assisted, while human contributions mainly involved directing, organizing, and verifying the AI’s work.

READ MORE: Anthropic launches Claude Opus 5 with major performance upgrades

In the second paper, researchers Milad Nasr and Nicholas Carlini examined shortened versions of the Advanced Encryption Standard (AES), the current global standard for securing internet data and communications. 

Cryptographers frequently examine these simplified versions to identify potential vulnerabilities in the overall system. However, the most common approach for analyzing such reduced versions has not evolved since 2013.

Working entirely on its own, the AI model invented a mathematical shortcut called a Möbius Bridge. This eliminates the trial-and-error phase that human researchers normally perform, speeding up testing by 200 to 800 times compared with previous methods. The researchers confirmed that this new system is straightforward to verify and audit.

The papers do not imply that existing encryption standards are vulnerable or that sensitive personal data has become suddenly insecure.

These two papers by Anthropic showed that AI can assist researchers in identifying hidden mathematical vulnerabilities and enhancing cryptographic systems before these issues become security threats, thereby strengthening defenses against future cyberattacks.