AI cloud infrastructure provider CoreWeave, Inc. has announced the official launch of its Sandboxes service. This offering provides AI researchers and platform teams with secure, isolated execution environments designed for workloads such as reinforcement learning, agent tool calling, and model evaluation.
Core Product Features Sandboxes functions as a unified execution layer that operates directly within a customer's CoreWeave Kubernetes Service cluster, eliminating the need for an independent execution stack. The service includes a Python SDK to support the creation and management of isolated, secure runtime environments. Each sandbox runs by default in a fully isolated virtual environment, ensuring that a failure or memory spike in one sandbox does not impact others.
Two Usage Models The Sandboxes service offers two access modes. For teams with an existing CoreWeave cluster, the service can run directly within that cluster. For teams without a dedicated cluster, a serverless runtime is available through Weights & Biases. Researchers can begin using the service within minutes by authenticating with their existing W&B API key and installing the Python client.
Industry Response Brian Belgodere, a senior technical member for AI/ML systems at IBM Research, stated that Sandboxes addresses a critical gap in their AI research stack by enabling secure, isolated, and large-scale code execution within existing compute resources. Roman Soletskyi, a scientist at Mistral AI, also confirmed that the service allows them to run hundreds of concurrent sandboxes with a single setup.
Comments