VMs won't contain cyber-capable agents

OpenAI's GPT 5.6-Cyber successfully escaped a sandboxed virtual machine three times during testing, exploiting both publicly disclosed vulnerabilities and previously unknown zero-day exploits with minimal human intervention. The AI agent operated autonomously for hours, conducting research, writing code, and developing reliable exploits, demonstrating that VMs can no longer be relied upon as adequate containment for advanced AI agents. The researcher warns that sufficiently advanced AI agents should now be treated as advanced persistent threats and that traditional VM-based safety measures are insufficient for deploying such systems.