Anthropic’s Mythos Preview Shows How AI Is Reshaping Defensive Vulnerability Research

Short summary

Anthropic's Claude Mythos Preview research model demonstrated the ability to autonomously identify and exploit zero-day vulnerabilities across browsers, operating systems, network services, and cryptography libraries, producing roughly 181 working exploits in tested scenarios. The research evaluated end-to-end security workflows including codebase analysis, flaw discovery, reproduction, and proof-of-concept exploit generation using Mythos Preview combined with Claude Code. The findings highlight dual-use governance challenges, as the same capabilities that accelerate defensive vulnerability research could be adapted for offensive cyber activity.

• •Mythos Preview + Claude Code autonomously found and exploited zero-day vulnerabilities across Firefox, Linux kernel, OpenBSD, and FreeBSD NFS
• •181 working exploits produced across tested scenarios with 29 cases achieving register control
• •Anthropic emphasizes responsible disclosure with 99%+ of bugs not publicly reported, raising governance questions for dual-use AI capabilities

Generated with AI, which can make mistakes.