Anthropic Ships Inference Hooks for Claude Enterprise: Real-Time DLP Before Claude Sees Your Data

Every Claude Enterprise prompt now optionally routes through your org's own security server — Netskope, Palo Alto, Zscaler, or a custom webhook — before the model ever processes it.

Launched August 5 in beta, inference hooks adds a compliance checkpoint at the inference layer. Turn it on once at the org level and it covers all Claude surfaces: chat, Claude Code, Cowork, and tool calls made through MCP connectors and skills — one configuration, all surfaces.

How it works: every inference request goes through a signed WebSocket connection to your DLP server, which returns an allow, modify, or block decision in milliseconds. The protocol uses an open, published schema — plug it into the DLP tool your security team already operates (Netskope, Palo Alto Networks, Proofpoint, Zscaler) or a custom endpoint you write yourself.

Three rollout options lower adoption friction: shadow mode (always allow, just log) for a zero-disruption audit pass; role-based exclusions to bypass specific users or groups; and percentage-based rollouts to canary before full deployment. Timeout tolerance and failure policy are configurable per org.

Why it matters: As agentic pipelines move into production — Claude Code reading sensitive codebases, Cowork sessions touching customer data — intercepting at the model boundary rather than auditing after the fact is a compliance prerequisite. Inference hooks delivers this without per-app code changes, which matters when your org runs Claude across dozens of surfaces simultaneously.

Sources: Inference hooks announcement — claude.com/blog · Anthropic puts inline DLP inside Claude Enterprise — Unite.AI · The Next Web coverage