Claude Enterprise can now ask your security team before it answers.

What changed

Anthropic launched inference hooks in beta for Claude Enterprise. Requests travel over a signed WebSocket to a data-loss-prevention server the customer controls, which returns allow or deny before inference begins. It covers chat, Claude Code, Cowork and tool calls through MCP connectors, skills and plugins, integrates with Netskope, Palo Alto Networks, Proofpoint and Zscaler or a custom server, and supports shadow mode, role exclusions and percentage rollouts.

Why it matters

Until now, stopping sensitive data reaching an AI model relied on training and hope. A checkpoint before inference turns policy into enforcement, but it also means somebody has to write the policy, and a badly tuned one silently blocks legitimate work or waves through what it should not.

Who should care

• Security and compliance teams governing enterprise AI use
• IT leaders rolling Claude out to teams handling regulated data

What to do

Start in shadow mode against real traffic and read what would have been blocked before enforcing anything. Assign an owner for false positives, and treat the hook as one control in a review chain, not proof that nothing sensitive gets through.