Anthropic has launched inference hooks, a beta feature for Claude Enterprise that routes every employee prompt through the organization’s own security server for an allow-or-deny verdict before the model sees it.
Launched August 5, the system covers claude.ai, Claude Code, and Claude Cowork with one organization-level configuration and nothing installed on user devices. When a user submits a prompt, Anthropic sends the transcript over a signed connection and waits, five seconds by default, for a verdict. Denied requests never reach the model, and every denial lands in the organization’s activity feed.
The same inspection runs on tool calls and their responses, including tools connected through MCP connectors, skills, and plugins, closing the path where sensitive data enters a conversation from connected systems. The security server sees transcript text, tool calls, and extracted attachment text, but never raw file bytes, system prompts, or tool definitions. Voice mode is not covered.
Anthropic is not selling a scanning engine. The webhook protocol has a published schema, and the company names Netskope, Palo Alto Networks, Proofpoint, and Zscaler as integration targets.
Rollout controls include shadow mode, percentage-based inspection, and role exclusions, with a configurable failure posture if the server is unreachable. Verdicts are binary for now, with response-side enforcement planned later.
