Anthropic Launches Inference Hooks, Letting Enterprises Screen Every Claude Prompt Before It Reaches the Model

Overview

Anthropic has launched inference hooks, a beta feature for Claude Enterprise that lets an organization’s compliance team inspect and approve every prompt before Claude processes it, according to Anthropic. The feature routes governed prompts and tool-call responses through a security server the customer or its security vendor operates, which returns an allow-or-deny verdict before inference runs, according to Anthropic’s documentation.

What We Know

When a user submits a prompt on a governed surface, Anthropic sends the conversation transcript to the organization’s configured security-server endpoint over an HTTPS POST request, with each request signed according to the Standard Webhooks specification so the receiving server can verify it came from Anthropic, according to Anthropic’s documentation. The security server has up to a configurable verdict timeout — five seconds by default — to respond, and Anthropic’s documentation states that “a denied request never reaches the model.”

Verdicts are strictly binary. As Anthropic’s documentation puts it, “Verdicts are allow or deny. Rewriting or redacting a prompt is not supported.” When a request is denied, the user sees a blocked-by-policy message that combines a per-request reason supplied by the security server with a standing message the organization’s administrators configure, and each denial is recorded in the organization’s Activity Feed, per the documentation.

The same inspection applies to tool calls. According to The Next Web, “When Claude calls tools through MCP, skills, or plugins, the tool’s response is checked before it reaches the model.” Anthropic’s documentation adds that the security server “sees what the user sees: transcript text, tool calls and their results, and text extracted from attachments,” but “never receives raw file or image bytes, system prompts, or Anthropic-internal context.”

One hook configuration covers Claude Enterprise activity across claude.ai chat, Claude Code, and Claude Cowork, whether accessed on the web, in the desktop app, or in the CLI, according to Anthropic’s documentation. The Next Web and Unite.AI both report that the feature is designed to plug into existing data-loss-prevention infrastructure, naming Netskope, Palo Alto Networks, Proofpoint, and Zscaler as compatible vendors alongside custom-built security servers.

Anthropic built rollout controls so organizations are not forced to block traffic immediately. Anthropic’s documentation describes shadow mode, which “observes verdicts on live traffic without blocking anything”; a rollout percentage that “inspects a chosen fraction of requests”; and exclusions that “exempt members of chosen roles entirely.” If the security server is unreachable, errors, or times out, an organization’s failure-handling setting determines whether the request is blocked or allowed through without inspection, per the documentation.

Currently, only one hook event exists. Anthropic’s documentation states that “today the only hook event is prompt, which fires once per governed inference request, before inference begins,” and that “response-side enforcement is planned as a later event.” The feature is not available for API access through the Claude Platform, nor on Amazon Bedrock or Google Cloud, and voice mode and ancillary requests such as conversation-title generation are not covered, according to the documentation.

Configuring inference hooks requires the organization:manage permission, held by an organization’s built-in Admin, Owner, and Primary owner roles, according to Anthropic’s documentation. Inference hooks is available in beta to Claude Enterprise organizations as of August 5, 2026, according to Anthropic and Unite.AI.

What We Don’t Know

Anthropic has not said when response-side enforcement — inspecting Claude’s output rather than only the incoming prompt — will become available, only that it is planned. Pricing details for inference hooks, and how many organizations have enabled it since launch, have not been disclosed in the documentation or in either outlet’s coverage reviewed for this article.

Analysis

Inference hooks formalizes a pattern that security teams have been building around large-language-model deployments piecemeal: an inline checkpoint that can veto a request before it reaches the model, rather than only auditing what already happened. Anthropic’s own comparison in its documentation frames the distinction directly — inference hooks act “inline, before inference runs,” while its existing Compliance API works “after the fact” to retrieve activity for audit and export, according to Anthropic’s documentation. By supporting the Standard Webhooks signing specification and citing compatibility with established DLP vendors, Anthropic is positioning the feature as a slot-in addition to security stacks enterprises already run, rather than a replacement for them.