Now loading...
Researchers at Anthropic have unveiled significant advancements in cryptography using their Claude Mythos Preview, revealing techniques that expose vulnerabilities in key cryptographic algorithms. The findings include a potent attack on HAWK, a digital signature scheme designed for resilience against quantum computing threats, as well as a novel approach for undermining reduced-round versions of the Advanced Encryption Standard (AES), one of the most widely adopted cryptographic ciphers. While these discoveries represent substantial progress, they do not currently impact operational systems.
The Claude Mythos Preview is capable of identifying and exploiting weaknesses in cryptographic libraries, which are essential for securing digital communications. Previously, the vulnerabilities identified were primarily due to flawed implementations of the algorithms. However, the recent findings indicate that Claude can also detect mathematical flaws within the algorithms themselves, potentially posing risks to billions of users relying on these systems for secure online activities.
The attack on HAWK represents a breakthrough, as this digital signature scheme is currently under consideration by the United States Government’s National Institute of Standards and Technology (NIST) due to its potential to remain secure in a post-quantum computing environment. Using Mythos Preview, researchers improved the most effective known attack on HAWK in just 60 hours, effectively reducing its key strength and raising concerns about its viability as a secure option.
In a related finding, the research team targeted reduced variants of AES, further investigating the security of this longstanding encryption standard. Mythos identified a method that enhances attack efficiency by 200 to 800 times, significantly accelerating the process of breaking a weaker 7-round version of AES, although these results do not compromise the full 10-round algorithm. Both of these discoveries were made in a context where practical implications for current systems remain limited.
One notable aspect of this research is Mythos’s ability to operate with minimal human intervention. An Anthropic researcher collaborated with Claude to develop the HAWK attack, while another facilitated Claude’s discovery of the AES breach through a semi-autonomous framework. The financial cost associated with these research efforts was approximately $100,000 per result, prompting further investigations resulting in additional attacks.
The research team is also focused on advancing the study of cryptography through collaborations with academic institutions and the creation of a benchmark known as CryptanalysisBench. This initiative aims to enhance the evaluation of language models like Claude Mythos in cryptographic contexts. The findings will be shared openly, including confirmation consultations with relevant stakeholders to ensure the accuracy and transparency of the results.
While the research underscores the immense potential for AI models to contribute to cryptographic advancements, it also brings to light the urgency for ongoing scrutiny of cryptographic standards as technology continues to evolve. With AI’s role in vulnerability discovery becoming more pronounced, researchers anticipate the need for a closer examination of the implications for existing cryptographic frameworks and the development of more secure algorithms for the future.
