Anthropic's Unreleased AI Model Chains Flaws Into Working Chrome, Safari, Windows and Linux Exploits

Anthropic says an internal, unreleased AI model discovered and chained multiple vulnerabilities into working proof-of-concept exploits for Chrome, Safari, Android, Windows and Linux, built entirely in a lab setting rather than observed in the wild. All issues were reported to the respective vendors first and are already fixed, including 62 Chrome bugs (critical/high severity) credited partly to Anthropic researcher Xinyang Ge "assisted by Claude," and an iOS/iPadOS Safari same-origin-policy bypass patched in version 26.5.2. The Windows chain, tied to CVE-2026-33096 (a denial-of-service flaw in HTTP.sys, CVSS 7.5), and the Linux chain, tied to CVE-2026-98432 (a 9.8-rated kernel use-after-free allowing root access), were both patched earlier this year, though independent CVE records don't fully match Anthropic's technical descriptions. The report matters because it suggests AI models can now autonomously combine separate flaws into functioning exploits—a task previously reserved for specialist human teams—though Anthropic hasn't disclosed exactly how much human guidance was involved or named the model used.