Anthropic launches inference hooks for inline data loss prevention in Claude Enterprise

Anthropic has introduced inference hooks, a new security feature for Claude Enterprise that enables inline data loss prevention (DLP) across all enterprise surfaces. Previously, inline enforcement was limited to Claude Code's client-side hooks; this update extends a single enforcement layer to chat, Claude Cowork, and other products without requiring separate integrations.

How it works

When enabled, every inference request routes through a signed WebSocket connection to a designated security server. Before generating a response, Claude sends the prompt and context to the server for inspection. The server returns an allow or deny decision, which Claude enforces before proceeding. This check also applies to tool calls, including those via MCP, skills, and plugins, ensuring that tool responses are vetted before reaching the model.

Integration and rollout

The feature uses a webhook-based protocol with a published schema, allowing organizations to integrate with existing DLP providers such as Netskope, Palo Alto Networks, Proofpoint, Zscaler, or custom in-house AI security servers. Rollout options include shadow mode (always allow), role-based exclusions, and percentage-based deployments to manage risk tolerance. Inference hooks is available today in beta for Claude Enterprise customers.