Anthropic announced its newest AI system, named Claude Mythos, on Tuesday. Described as a versatile, unreleased advanced model, it demonstrates such significant capabilities that the company hesitates to distribute it widely to the public.
The preview version of Claude Mythos exceeds the performance of nearly all human experts in detecting and taking advantage of software weaknesses. Recent demonstrations show it independently identified thousands of unknown zero-day vulnerabilities over the last few weeks.
By uncovering these security issues on its own, Claude Mythos Preview highlights its potential importance in enhancing cybersecurity practices moving forward. Anthropic aims to direct its application toward positive outcomes, such as countering the growing dangers posed by widespread use of generative AI technologies.
We have no intention of offering Claude Mythos Preview for broad access, though our long-term objective involves allowing users to implement models of this caliber securely on a large scale. Achieving that requires advancements in cybersecurity measures and similar protections to identify and prevent the system's most hazardous results.
Claude Mythos forms a key element of Project Glasswing, an extensive security program that includes collaboration with 11 partners: Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, Nvidia, and Palo Alto Networks.