Claude API adds IP allowlisting to blunt leaked keys
A beta feature in the Claude Console rejects API requests from outside a list of IP ranges you set, with up to 1,000 entries per organisation.
Anthropic added IP allowlisting to the Claude Platform on 5 November, in beta. An admin or owner sets a list of IP ranges in the Console, and Anthropic checks every covered request’s public IP against it, rejecting anything outside the list with a 403 permission_error.
The allowlist covers the Messages, Message Batches, Token Counting, Files and Skills API endpoints, plus the Admin and Compliance APIs, whether the caller uses an API key or Workload Identity Federation. Entries can be a single IPv4 or IPv6 address or a CIDR range, no broader than /8 for IPv4 or /32 for IPv6, up to 1,000 per organisation. Changes take up to 15 minutes to apply.
203.0.113.0/24
2001:db8::/32Why it matters: if your CI, servers or proxies have a stable set of outbound addresses, this makes a leaked API key far less useful to whoever finds it, since the key still won’t work outside your listed ranges. Anthropic is explicit that it limits where a key works, not how long, so it’s a complement to key expiration, not a substitute for it.
The caveat: it doesn’t apply to Claude in Amazon Bedrock, Google Cloud or Microsoft Foundry, and it is beta, so behaviour can change. List every network, including CI, before turning it on, or you can lock yourself out.