Discovering and exploiting vulnerabilities in real-world software with Claude

Anthropic’s Frontier Red Team used an unreleased model to discover and build five exploit chains in widely used browsers and operating systems, often with little or no user interaction. In Chrome, the model combined flaws to escape the browser sandbox and run code as the user; Google assigned 62 CVEs, all rated high or critical, and patched them. In Safari, a memory-corruption flaw and weaknesses in protections let a website bypass the same-origin policy and access sensitive data without executing code. An Android chain bypassed a locked phone’s USB and pairing safeguards to extract personal files. The model also found a remotely exploitable Windows kernel flaw that could repeatedly crash servers, and a Linux kernel race condition that enabled unauthenticated remote command execution as root. The affected vendors addressed the reported vulnerabilities.

The team’s main research finding is that running many varied scanning agents in parallel, with more computation at test time, substantially increases the number of vulnerabilities found. Their approaches ranged from checking individual functions to testing whether software violated its security assumptions; confirmed findings also underwent verification, risk rating, and automated searches for related flaws. They argue that patching known bugs is only half the task: developers should also scan code changes as they are proposed to prevent new vulnerabilities. The exploits demonstrate that models can combine multiple weaknesses into working attacks - work once limited to specialist teams - so defenders need access to similarly capable tools and security checks integrated into development.