Anthropic launches inference hooks: inline data loss prevention for Claude Enterprise, including Cowork

Anthropic launched inference hooks on August 5, 2026, in beta for Claude Enterprise. The idea is simple and, for compliance teams, long-awaited: before a prompt reaches the Claude model, Anthropic first sends the conversation to an AI security server that your organization (or its security vendor) operates over HTTPS. That server returns an allow or deny verdict. A denied request never reaches the model at all.

The checkpoint moves to the server side. Traditional DLP for AI tools relies on browser plugins or network proxies on every device. Inference hooks run on Anthropic's side, after a request leaves the client and before inference starts. One organization-level configuration covers every governed surface, chat, Claude Code, and Claude Cowork included, with nothing to install on user devices.

Tool calls are inspected too. The same check runs when Claude calls a tool through MCP connectors, skills, or plugins: the tool's response is checked before it reaches the model. That matters for Cowork specifically, where a single session can touch mail, drives, CRM systems, and internal databases through connectors.

Built to plug into existing security stacks. The protocol is webhook-based with a published schema, designed to work with DLP infrastructure from Netskope, Palo Alto Networks, Proofpoint, and Zscaler, or a custom-built server. Rollout controls are pragmatic: a shadow mode (always allow, but log) lets teams monitor before enforcing, and role-based exclusions plus percentage-based rollouts support a gradual introduction.

The honest limitations. Verdicts are binary: a server can allow or deny a prompt, but cannot rewrite or redact it. And at launch the only hook event fires on the prompt before inference; response-side enforcement, checking what the model returns, is planned as a later event. It is also Enterprise-only for now.

Why this matters for Cowork organizations. Until now, the standing compliance caveat around Cowork was that its activity is hard to govern centrally. Inference hooks are the first inline, organization-wide control that includes Cowork sessions, a meaningful step for regulated industries evaluating agentic AI.

Also this week in brief: Anthropic extended the 50% weekly usage boost for Claude Code through August 19 (Pro, Max, Team, and seat-based Enterprise), and the introductory API pricing for Sonnet 5 ($2/$10 per million tokens) ends August 31; standard pricing of $3/$15 applies from September 1.