A New Checkpoint Across Claude Enterprise! Inference Hooks Now Blocks Data Leaks in Real Time

Hi, it's Shii! Today I found news that should make security teams smile. Claude for enterprises just got a new mechanism to stop sensitive data from leaking out before it happens. Let me walk you through it!

Claude Blog claude.com

What was announced?

Anthropic's Claude Blog announced a new feature for Claude Enterprise called Inference Hooks. It lets a company's compliance team inspect every prompt and every tool-call response before it reaches Claude, and enforce policy by approving or blocking it. It applies across all Claude Enterprise surfaces, including chat, Claude Code, and Claude Cowork.

The division of labor is simple: your organization's DLP (data loss prevention) server decides whether something should pass or get blocked, and Claude enforces that decision in real time.

The story so far

Until now, Claude's native inline enforcement was limited to Claude Code's client-side hooks. That meant separate integration work for each product, and no way to meet a security team's requirement that every channel where employees can move sensitive data passes through an inspection point they control, across all of Claude Enterprise at once.

What changes

Inference Hooks is positioned as a single enforcement layer that covers Claude Enterprise surfaces without needing a separate agent or integration for each product. Here's what it covers:

• Chat
• Claude Code
• Claude Cowork
• Tool calls made through MCP connectors, skills, and plugins

Turn it on once at the organization level, and it applies everywhere, so you don't need to configure each product separately. Andrew Grimmett, Vice President of Information Security at Bandwidth, one of the companies already using it, put it this way:

Inference hooks add a checkpoint to inspect what's flowing to Claude in real time, before anything sensitive leaves our environment. This lets us safely move faster on AI without giving up control.

Dive Deep

Let's look a bit closer at how this actually works.

The request flow: When an organization turns on Inference Hooks, every inference request routes through a signed WebSocket connection to the organization's security server. Before the model starts generating, Claude sends the prompt and its surrounding context to that server, and waits for an allow or deny verdict before proceeding. Tool-call responses go through the same check before being sent back to the model.

Extending your existing DLP setup: Inference Hooks runs on an open, webhook-based protocol with a published schema, so you can point it at the same server your other tools already report to. Vendors mentioned as compatible include:

• Netskope
• Palo Alto Networks
• Proofpoint
• Zscaler
• A custom in-house AI security server

Rolling it out safely: You don't have to flip this on for the whole company at once. Available options include:

• Shadow mode (always allow, so you can observe first)
• Role-based exclusions
• Percentage-based, staged rollouts

You can also customize the failure policy for when the verdict server doesn't respond, along with timeouts, to match your organization's risk tolerance.

Availability: Inference Hooks is available today in beta for Claude Enterprise customers. For security vendors, since it's built on a published webhook-based protocol and schema, you can build your own integration so that Claude Enterprise customers can point their organization at your platform.

Wrap-up

Today I introduced Inference Hooks, Anthropic's new feature for Claude Enterprise. Here's the recap:

• An organization's DLP server can inspect every prompt and tool-call response in real time before it reaches Claude, and approve or block it
• It's a single enforcement layer covering chat, Claude Code, Claude Cowork, and tool calls through MCP connectors, so there's no need to integrate each product separately
• It runs over a signed WebSocket connection and a published webhook protocol, working with existing DLP infrastructure like Netskope, Palo Alto Networks, Proofpoint, Zscaler, or your own custom server
• Shadow mode, role-based exclusions, and percentage-based rollouts let you deploy it safely in stages
• It's currently in beta for Claude Enterprise customers

If you're on a security or compliance team trying to get a unified view and control over every channel where employees use Claude, this one's especially for you!