Claude Enterprise Now Asks Your Server for Permission. You Get 5 Seconds.
Anthropic's inference hooks put your DLP server in front of every prompt, with a five-second default timeout. Governance just became a blocking dependency in the request path — and an availability SLO you now own.
NeuroX AI · August 6, 2026
Anthropic shipped inference hooks on August 5 — beta, Claude Enterprise. Every prompt routes through your own security server for an allow-or-deny verdict before the model generates anything. Your server gets five seconds by default to answer.
The inspection surface is specific. Your server receives transcript text, tool calls and their results, and text extracted from attachments — not raw file bytes, not system prompts, not tool definitions. It covers chat, Claude Code, and Cowork, including tool calls made through MCP connectors, skills, and plugins. One org-level config, a signed WebSocket, nothing installed on laptops.
Read the timeout again, because that's the actual news. Enterprise DLP has always been advisory — it watched traffic and filed alerts. This puts it in the request path as a blocking call. When your DLP server is slow or down, you pick: block the request and every engineer's Claude Code session stalls, or let it through uninspected and the control you deployed isn't one. Anthropic makes that posture configurable instead of choosing for you, which is the honest design.
So you now own an availability SLO for a service most security teams have never run hot. That's why shadow mode, percentage rollouts, and role exclusions ship in the same release. Treat it like blocking middleware, not a policy toggle.
One gap to check before you plan around it: this covers Enterprise surfaces only. Agents running against the API, Bedrock, or Vertex sit outside it.