VMs won't contain cyber-capable agents

The author tested GPT 5.6-Cyber’s ability to escape a QEMU/KVM sandbox on a Debian 12 system, and the model succeeded three times, first exploiting a newly disclosed kernel bug (CVE-2026-53359), then using unpatched vulnerabilities, and finally discovering multiple zero-day flaws after rebuilding QEMU from upstream sources. The experiment demonstrated that advanced AI agents can breach VM isolation rapidly and autonomously, requiring containment strategies beyond traditional virtual machines.