Claude Fable 5 Tests smolvm for Untrusted Code

Recent testing of smolvm version 1.8.3 has demonstrated its viability as a secure, hardware-isolated sandbox for executing untrusted Python and JavaScript data transformations. Unlike traditional shared-kernel containers, smolvm leverages hardware-isolated virtual machines to enforce strict security boundaries. The evaluation revealed highly competitive performance metrics, with cold start times ranging between 0.6 and 1.5 seconds, while warm executions clocked in at a rapid 50 milliseconds.

During the testing process, several critical security and resource-management features performed exactly as intended. These included offline local images, execution without network access, strict CPU and RAM limits, and guest-enforced timeouts. Additionally, the system successfully handled storage quotas, read-only input mounts, writable output mounts, and the --unprivileged flag, making it a robust option for developers who need to run user-provided code safely.

The evaluation itself highlighted the advanced capabilities of modern AI agents. Researcher Simon Willison tasked Claude Fable 5, running within the Claude Code for web environment, to conduct the research. The AI encountered an immediate roadblock because the Claude Code container, running Linux 6.18.5-fc-v20 as a Firecracker guest with 4 vCPUs and 15GB of RAM, lacked KVM access, preventing nested virtualization. To bypass this, the AI autonomously pivoted to running the test suite via a temporary GitHub Actions workflow on an Ubuntu runner, which does expose KVM.

For software practitioners, these findings confirm that smolvm offers a practical, high-performance alternative to container-based sandboxing. By utilizing hardware isolation rather than shared kernels, developers can mitigate the risks of container escape vulnerabilities when processing untrusted user scripts. The integration of such tools, especially when validated by proactive AI agents, points to a future of automated, highly secure deployment pipelines.