Vercel Sandbox can now enforce egress network policies through Server Name Indication (SNI) filtering and CIDR blocks, giving you control over which hosts a sandbox can reach. Outbound TLS connections are matched against your policy at the handshake, unauthorized destinations are rejected before any data is transmitted.

By default, sandboxes have unrestricted internet access. When running untrusted or AI generated code, you can lock down the network to only the services your workload actually needs. A compromised or hallucinated code snippet cannot exfiltrate data or make unintended API calls, traffic to any domain not on your allowlist is blocked.

Link to headingGoing beyond IP based rules to host based

The modern internet runs on hostnames, not IP addresses, a handful of addresses serve thousands of domains. Traditional IP-based firewall rules can't precisely distinguish between them.

Host-based egress control typically requires an HTTP proxy, but that breaks non-HTTP protocols like Redis and Postgres. Instead, we built an SNI-peeking firewall that inspects the initial unencrypted bytes of a TLS handshake to extract the target hostname. Since nearly all internet traffic is TLS-encrypted today, this covers all relevant cases. For legacy or non-TLS systems, we do also support IP/CIDR-based rules as a fallback.

Link to headingRestrict to specific hosts at creation

Define which domains the sandbox can reach. Everything else is denied by default. Wildcard support makes it easy to allowlist services behind CDNs:

import { Sandbox } from '@vercel/sandbox';
const sandbox = await Sandbox.create({
  networkPolicy: {
    allow: ['ai-gateway.vercel.sh', '*.vercel.com'],
  },
});
// Can reach Vercel AI Gateway and Vercel, nothing else
await sandbox.runCommand('node', ['process-data.js']);

Link to headingAdjust after initial setup

Policies can be updated dynamically on a running sandbox without restarting the process. Start with full internet access to install dependencies, lock it down before executing untrusted code, reopen to stream results after user approval, and then air gap again with deny-all, fully in one session:

import { Sandbox } from '@vercel/sandbox';
const sandbox = await Sandbox.create();
// Phase 1: Open network, download everything we need
await sandbox.runCommand('npm', ['install']);
await sandbox.runCommand('aws', ['s3', 'cp', 's3://my-bucket/input-data', './data', '--recursive']);
// Phase 2: Lock down, only the AI gateway is reachable
await sandbox.updateNetworkPolicy({
  allow: ['ai-gateway.vercel.sh'],
});
// Run the untrusted / agent-driven workload in isolation
await sandbox.runCommand('node', ['agent.js']);
// Phase 3: Open a narrow hole to post results back
await sandbox.updateNetworkPolicy({
  allow: ['my-bucket.s3.amazonaws.com'],
});
await sandbox.runCommand('aws', ['s3', 'cp', './output/results.json', 's3://my-bucket/output/results.json']);
// Phase 4: Lockdown Internet access
await sandbox.updateNetworkPolicy('deny-all');

Read the documentation to learn more about network egress firewall policies, available on all plans.