Google launches Cloud Run sandboxes to isolate and execute untrusted AI-generated code at millisecond speed

Google announced the public preview of Cloud Run sandboxes, a service that isolates and executes code or prompts generated by AI that cannot be trusted, within the existing Cloud Run environment without additional cost. The feature is designed to run untrusted code or prompts safely, preventing malicious actions from affecting the host system or other workloads. Cloud Run sandboxes provide a sandboxed execution environment for AI-generated code, such as Python scripts created by large language models, ensuring complete isolation from the host system. This allows tasks like processing sensitive data or running untrusted web requests without exposing the broader system to risks. The service is intended for use cases including data analysis, web scraping, and automated workflows, where code execution must be secure and isolated. Key capabilities highlighted include the ability to execute up to 1,000 sandbox instances in a single Cloud Run deployment, with processing times for non-malicious Python code reportedly reduced to around 500 milliseconds. The sandbox environment enforces strict isolation, preventing data leakage and unauthorized network access, and automatically terminates after execution to avoid lingering resources. Google identifies three primary use cases for Cloud Run sandboxes. First, it supports AI code interpretation, enabling high-speed data processing and execution of Python or R scripts for tasks such as model inference, data partitioning, and metric generation. Second, it facilitates web request handling, allowing secure execution of webhooks and browser automation without exposing the host system. Third, it enables developers to run code uploaded by users directly on Cloud Run, supporting platform integrations, plugins, and automated workflows. The service integrates with the Agent Development Kit (ADK) under the Agent Development Kit framework, where a single line of code can add sandbox execution capabilities to AI agent workflows. It also extends the Compute SDK, enabling sandbox execution from external systems or within the same service region. Cloud Run sandboxes are positioned as a cost-effective alternative to traditional virtual machines, offering faster startup and execution while maintaining strong isolation and security. The public preview of Cloud Run sandboxes is scheduled for release on July 9, 2026, in Japan Standard Time.

Reported by ITmedia AtIT.