Anthropic's Frontier Red Team says an unreleased research model found flaws in Chrome, Safari, Android, Windows and Linux and chained them into proof-of-concept exploits, needing only "initial prompting" and little human involvement. The exploits were built in a lab and have not been seen in real attacks; all flaws were reported to vendors.
The Chrome chain strings together four bugs to escape the browser sandbox from a single page visit; Anthropic reported 62 critical or high flaws to Google, 13 of them sandbox escapes, fixed in Chrome 156 on October 20. The Linux chain leaks a kernel address and reaches a remote root shell without login (CVE-2026-98432, CVSS 9.8, fixed October 15); Anthropic says the leak took 48 minutes and the shell five more. The Safari chain lets a malicious site read data from other sites a user is signed in to (fixed in iOS 26.5.2), while the Windows chain only crashes a server (CVE-2026-33096).
The Hacker News notes that the model is unnamed, Anthropic does not break down the human share of the work, and some CVE links rest mainly on Anthropic's own account. Users should install the vendor patches.