Anthropic experiment shows unreleased model can automatically turn Linux kernel vulnerability into remote root exploit chain

Anthropic disclosed that in experiments, an unreleased frontier model successfully turned a theoretical Linux kernel vulnerability with a CVSS score of 9.8 into an unauthenticated remote root exploit chain. Tested against CVE-2026-98432—a use-after-free flaw in the IPv4 TCP handshake—the model completed a kernel pointer leak and gained a root shell in about 53 minutes, drastically compressing an exploit weaponization process that typically takes weeks.