Discovering and exploiting vulnerabilities in critical software with Claude

An Anthropic red team used an unreleased Claude model to find and demonstrate serious vulnerabilities in widely used browsers and operating systems. Its five exploit chains included Chrome flaws that enabled one-click code execution beyond the browser sandbox; a Safari chain that bypassed same-origin protections and exposed sensitive browsing data without executing code; an Android lock-screen bypass that extracted files from a locked phone; a remotely triggered Windows kernel flaw that could repeatedly crash servers; and a Linux kernel bug that allowed unauthenticated, zero-click command execution as root. The Linux exploit took about 53 minutes to achieve its two stages: leaking kernel memory information and hijacking control flow. The team reported the issues to vendors; the described flaws have been patched, including 62 Chrome vulnerabilities rated high or critical.

The work argues that models can now combine multiple weaknesses into working exploits with limited human involvement, a capability once concentrated among specialist security teams. The team increased discovery by running diverse scanning agents in parallel, varying their goals and threat models, and using automated analysis to find related flaws. Standard verification and risk-rating processes helped maintain confidence in findings. It recommends both clearing existing vulnerability backlogs and preventing new flaws through security scanning during code review. The central implication is that defenders need access to similarly capable models and development tools to keep pace with automated vulnerability discovery and exploitation.