Skip to content

Claude Enterprise inference hooks reach general availability

Admins can now route every prompt through their own security server for a real-time allow or deny verdict before the model sees it, with a 5-second default timeout and DLP as the main use case.

Source: Claude Platform docs

Anthropic took Inference hooks out of beta for Claude Enterprise organisations on 20 October. The feature lets a security team point Claude at their own HTTPS “AI security server”: every governed prompt, and every tool call if Validate tool calls is on, gets sent there for an allow or deny verdict before inference runs.

Requests are signed to the Standard Webhooks specification, and the server gets a configurable timeout, 5 seconds by default, to respond. A deny blocks the request before it reaches the model and logs it to the compliance Activity Feed. A sustained outage on your side trips a circuit breaker, so you are not left silently blocking or allowing everything. One hook covers claude.ai, Cowork, Claude Code, Claude for Microsoft 365, Claude Tag and Claude in Chrome, though not Amazon Bedrock or Google Cloud deployments.

POST https://your-security-server/verdict
{"action": "allow"}

Why it matters: this is inline, real-time control over what reaches the model, rather than after-the-fact auditing through the Compliance API. For teams with data loss prevention or model allowlisting requirements, enforcement can now live in infrastructure they already run instead of being bolted onto each client app.

The caveat: raw file and image bytes never reach your security server, so image-only content such as a screenshot is not inspected, and Platform (API) organisations are out of scope. Enterprise only.

Share

Related reading

Engineering

How to patch PHP in a hurry without breaking production

PHP ships an out-of-band fix for two critical flaws at 14:00 UTC today. Here is the hour of preparation that turns the patch into a non-event, and the runbook for when it lands.

ML for Web Devs

Tokens, context windows, and why your bill is what it is

The unit of billing for language models is not the request or the word. Once you understand tokens and how a context window fills up, the invoice stops being a surprise.

ML for Web Devs

RAG is just search with extra steps

Retrieval-augmented generation gets written up like an architecture. It is a search query, a string concatenation and one API call, and the search query is the part that decides whether it works.

More from The Wire

All briefs