Vercel AI Gateway now enforces a team-wide provider allowlist, blocking all traffic to unapproved AI vendors at the gateway level. This applies to every request format the gateway supports, including AI SDK, OpenAI Chat Completions, and Anthropic Messages API, and it covers BYOK traffic. Only team owners can modify the list, and new providers added to AI Gateway are disabled by default once the allowlist is active, so the approved set cannot silently expand.
The enforcement model matters more than the feature itself. Developers cannot route around it, and coding agents cannot either, even if they omit or modify request-level provider filters. The allowlist compounds with existing restrictions like Zero Data Retention and request-level filtering, operating as a logical AND across all active policies. Filtering is by provider, not by model, and the gateway will fall back to other allowed providers serving the same model if the primary fails.
The documentation details how the allowlist interacts with ZDR and prompt training restrictions, which is worth reading for teams managing layered compliance requirements. Configuration is in the AI Gateway Settings tab under Provider Allowlist. All providers are allowed by default, so existing traffic is unaffected until you start disabling vendors.
[READ ORIGINAL →]