Anthropic demonstrated that its AI-developed exploit for CVE-2026-98432 took approximately 53 minutes to achieve remote command execution by leaking kernel memory addresses before issuing shell commands