Vercel Sandbox Firewall Now Supports Outbound Request Proxying and Filtering.

Vercel released an update to its Sandbox firewall that enables proxying and filtering of outbound requests, giving developers more control over how agents and services communicate with external systems.

The update addresses a growing need as coding agents increasingly build and deploy software autonomously. According to Vercel's recent data, over 30% of deployments on the platform are now initiated by coding agents, up 1000% from six months ago. As these agents write code that calls external APIs and AI providers, managing outbound traffic becomes critical for security and compliance.

The firewall upgrade introduces three key features. First, developers can now route all agent traffic through their own proxy, allowing centralized control over which external services sandboxed code can reach. Second, credential injection has become more granular, letting teams scope API keys and tokens to specific requests rather than exposing them broadly to the sandbox environment. Third, Vercel added logging and audit trails for outgoing traffic, making it possible to track which external endpoints agents are calling and when.

These capabilities matter most for teams running AI agents that autonomously generate and execute code. Without outbound filtering, a sandbox could potentially call any public endpoint. With this update, developers can whitelist specific domains, inject credentials only when needed, and maintain records of external dependencies.

Vercel has been building what it calls "agentic infrastructure" to support the shift toward machine-driven software development. The company positions the update as one piece of a larger stack that includes Workflows for long-running processes, AI Gateway for model routing, and sandboxed execution environments.

The firewall upgrade complements custom tags for sandboxes, which shipped earlier this year to help teams organize and manage ephemeral compute environments at scale.

Vercel projects deployed by coding agents are 20 times more likely to call AI inference providers than those deployed by humans, according to the company's blog. Tools like Claude, Cursor, Lovable, and v0 account for the majority of agent-driven deployments on the platform.

The outbound request filtering feature is available now for Vercel Sandbox users. Full technical documentation is linked in the original announcement. The update ships alongside other governance improvements, including Trusted Sources for Deployment Protection, which lets teams authorize OIDC tokens from external services instead of sharing long-lived secrets.

Developers can configure proxy rules, credential scoping, and logging through the Vercel dashboard or API. The feature works with both Vercel's native OIDC tokens and external identity providers like GitHub Actions.