Anthropic disclosed that an Claude AI research model developed a working exploit for CVE-2026-98432 that gains root privileges on Linux servers