Novita AI, a San Francisco-based AI and agent cloud platform, has launched Novita Sandbox, a secure runtime infrastructure layer designed to isolate autonomous agent systems during execution. The product targets a growing class of security risks introduced by autonomous agents that execute code, call APIs, browse the web, and interact with live environments, including frameworks such as OpenClaw and Hermes Agent.
Each agent task in Novita Sandbox runs inside a dedicated Firecracker microVM with its own kernel, isolated memory boundary, and ephemeral filesystem, ensuring no shared process space, no credential leakage, and no cross-agent interference. The isolation is designed to eliminate exposure risks including access to developer machines, local files, API keys, cloud configurations, and active browser sessions. The company describes this as a new product category it calls Secure Agent Infrastructure.
The platform delivers sub-200 millisecond average startup times and scales to thousands of parallel microVMs under concurrent workloads, making it suitable for real-time and production-scale deployments. For long-running workflows, Novita Sandbox introduces stateful execution, allowing environments to be paused and resumed with full runtime state, including memory, filesystem, and active processes, restored in approximately one second without compromising isolation. Novita Sandbox is available immediately through managed NovitaClaw deployments, Sandbox Skills for OpenClaw, Hermes-compatible runtimes, and persistent node configurations. Novita AI’s platform supports fast inference across more than 120 large language and multimodal models through a single API.
KEY QUOTE:
“With this launch, Novita introduces a new category: Secure Agent Infrastructure. A trusted execution layer for deploying autonomous systems safely at scale without exposing local machines or sensitive runtime environments.”
Junyu Huang, COO, Novita AI
