Anthropic's internal Claude research model builds five exploit chains across Chrome, Safari, Android, Windows and Linux
Anthropic published a new vulnerability research report describing how an as-yet unpublished internal Claude research model autonomously located vulnerabilities and wrote exploit code to build five runnable attack chains across Chrome, Safari, Android, Windows and Linux. The Linux chain corresponds to CVE-2026-98432, where an attacker only needs to send packets to obtain root without credentials; Anthropic's verification took 53 minutes, 48 of them to leak kernel pointers and 5 to execute commands. The Android chain targets a locked Pixel 9 and extracts photos and other personal data over a USB-related flaw without unlocking, though physical access is required. The Windows chain maps to CVE-2026-33096 with a CVSS score of 7.5 requiring no authentication or user interaction. All vulnerabilities were reportedly fixed through coordinated disclosure.