Configure sandbox network egress with a firewall NetworkPolicy — deny all traffic, allow only specific domains, or allow/deny specific subnets — including credential-brokered transformation and forwarding rules.

Documented steps

1. Set the base mode at creation: `Sandbox.create({ networkPolicy: 'allow-all' })` (default, permits all egress) or `Sandbox.create({ networkPolicy: 'deny-all' })` (blocks all egress including DNS resolution). Docs: https://vercel.com/docs/sandbox/sdk-reference#networkpolicy-class
2. Restrict to specific domains: `networkPolicy: { allow: ['ai-gateway.vercel.sh', 'google.com'] }`. Matching is SNI-based (TLS only): exact match unless the domain uses a wildcard segment (`www.*.com` matches one segment; `*.google.com` matches any subdomain but not the bare `google.com`).
3. Allow/deny specific IP ranges alongside domain rules: `networkPolicy: { allow: [...], subnets: { allow: ['10.0.0.0/8'] } }` or `{ subnets: { deny: ['10.0.0.0/8'] } }` — `subnets.deny` always takes precedence over `subnets.allow` and domain-based `allow` entries.
4. To inject credentials into brokered requests without exposing secrets to the sandbox, define a `transform` rule with headers: `{ allow: { 'ai-gateway.vercel.sh': [{ transform: [{ headers: { 'x-api-key': 'secret-key' } }] }] } }` (Pro/Enterprise only — this intercepts TLS to alter requests).
5. Update an already-running sandbox's policy: `await sandbox.update({ networkPolicy: 'deny-all' })` (applies to current and future sessions); the older `sandbox.updateNetworkPolicy()` method still works but is deprecated in favor of `update()`.
6. From the CLI, use flags on `sandbox create`: `--network-policy deny-all`, `--allowed-domain <domain>`, `--allowed-cidr <cidr>`, `--denied-cidr <cidr>`.