Cloudflare says Anthropic's Mythos Preview finds exploit chains that earlier frontier models missed

chains that earlier frontier models missed">

In a significant advancement for cybersecurity, Cloudflare has tested Anthropic's security-focused AI model Mythos Preview across more than 50 of its own code repositories as part of Project Glasswing. The results show that Mythos Preview can chain multiple small vulnerabilities into working exploits, providing concrete proof by writing, compiling, and running proof-of-concept code autonomously. Earlier frontier models were able to identify similar individual bugs and sometimes provided solid analysis, according to Cloudflare's Chief Security Officer, Grant Bourzikas.

However, these models fell short when it came to connecting the dots, often leaving exploit chains incomplete and the question of actual exploitability unanswered. In contrast, Mythos Preview produced fewer speculative findings and provided clearer steps to reproduce issues, requiring less human follow-up to reach a fix-or-dismiss decision. Cloudflare emphasizes that relying on a single agent is not sufficient.

To enhance the accuracy and reliability of its findings, the company developed a multi-stage harness that utilizes up to 50 parallel agents and includes adversarial review. In this setup, a second agent attempts to disprove each finding, adding an additional layer of scrutiny. While this technology marks a significant leap forward in cybersecurity, Cloudflare also warns that these advanced capabilities will soon be available to attackers as well.

This dual-edged nature of the technology underscores the need for continuous innovation and vigilance in the field of cybersecurity. The testing and implementation of Mythos Preview by Cloudflare highlight the evolving landscape of cybersecurity, where AI and machine learning are increasingly playing critical roles. As these technologies continue to advance, their impact on both defensive and offensive cybersecurity strategies will be profound.