Anthropic used Claude to discover and exploit vulnerabilities in real-world critical software
Discovering and exploiting vulnerabilities in critical, real-world software with Claude
Anthropic used an unreleased internal research model to discover serious vulnerabilities in Chrome, Safari, Android, Windows and Linux, write five attack chains requiring little or no user interaction that achieve effects such as remote code execution, and report them to vendors under coordinated disclosure. The Chrome one-click RCE chain chained together vulnerabilities including a V8 heap corruption, a sandbox escape and a FedCM use-after-free; 62 CVEs were fixed in Chrome 156, all critical or high severity. The Safari chain bypassed the same-origin policy to read sensitive data such as cookies, and Apple confirmed and fixed it the same day; the Android lock-screen bypass could extract personal information such as photos without unlocking the device. The Windows kernel QUIC out-of-bounds read (CVE-2026-33096) and the Linux kernel multipath TCP race condition vulnerability (CVE-2026-98432) have both been fixed.
The original source gives the complete process of five chains from initial prompt to usable exploit chain and the CVE fix status; defense teams can refer to its parallel scanning and PR-stage scanning methods.