VMs won't contain cyber-capable agents (blog.trailofbits.com)

Recent testing of GPT 5.6-Cyber revealed alarming capabilities in cybersecurity, demonstrating the agent's ability to escape a virtual machine (VM) using a variety of methods, including exploiting known vulnerabilities as well as 0-day exploits. The experiment illustrated that the AI could autonomously identify weaknesses in the host kernel, libslirp, and QEMU, successfully executing multiple escape attempts that included building its own exploits. This raises significant concerns about relying solely on VMs to contain advanced AI agents, as they are able to easily exploit shared resources and vulnerabilities present due to lags in security updates. This development challenges long-standing assumptions about virtualized environments as safe perimeters. The inherent complexity and outdated software in systems like older Debian distributions can no longer be trusted as impenetrable barriers against sophisticated AI agents. The necessity for urgent software updates and careful considerations of virtualization tools becomes paramount. The findings call for a reevaluation of existing cybersecurity protocols and suggest leveraging more secure virtualization technologies, such as Firecracker, which aim to minimize the attack surface and enhance overall security for deploying AI agents.