Anthropic adds Inference hooks for Claude Enterprise

Anthropic announced Inference hooks on August 5, 2026, a beta control for Claude Enterprise organizations that lets an organization’s AI security server allow or deny governed requests before Claude runs inference.

How the policy gate works

When a user submits a governed prompt, Anthropic sends the conversation transcript to the organization’s configured HTTPS security server. The server returns an allow or deny verdict, and a denied request does not reach the model. Anthropic says requests are signed so the receiving service can verify their origin.

The official documentation says the current hook event is prompt, which fires before inference begins. A default verdict timeout is five seconds, and organizations can choose whether an unreachable or failing security server should block the request or allow it to proceed without inspection.

One organization-level setup

Anthropic says one hook can govern conversations across claude.ai, Cowork, and Claude Code, including web, desktop, and CLI sessions in the organization. Common uses include data-loss prevention, real-time transcript archival, prompt telemetry, and policy engines for model or project restrictions.

The security server sees transcript text, tool calls and their results, and text extracted from attachments. It does not receive raw file or image bytes, system prompts, or Anthropic-internal context. Image-only content is therefore not inspected through the extracted-text path described in the docs.

Another August 5 model change

Anthropic’s same-day API release notes also say the Claude Opus 4.1 model (claude-opus-4-1-20250805) has been retired. Requests to that model now return an error, and Anthropic recommends upgrading to Claude Opus 5; researchers can request ongoing access through the External Researcher Access Program. If a deployment still pins the Opus 4.1 identifier, this is a migration item separate from the new governance hook.

Availability and limits

Inference hooks are available in beta to Claude Enterprise organizations. Configuration requires the organization:manage permission. Anthropic’s documentation says the feature is not available on Amazon Bedrock or Google Cloud, and that platform organizations using Claude API access are out of scope.

The release is different from the client-side hooks developers configure inside Claude Code: Inference hooks run in Anthropic’s infrastructure before inference and apply across the organization’s covered Claude surfaces. Teams should still review the beta protocol, failure mode, and data-handling details before using it for production policy enforcement.

Source

• Anthropic — “Inference hooks: inline data loss prevention for Claude Enterprise” (August 5, 2026): https://claude.com/blog/claude-enterprise-inference-hooks
• Claude Platform Docs — “Inference hooks”: https://platform.claude.com/docs/en/manage-claude/inference-hooks
• Claude Platform — API release notes (August 5, 2026): https://docs.anthropic.com/en/release-notes/api
• Claude Platform Docs — “Develop an Inference hooks integration”: https://platform.claude.com/docs/en/manage-claude/inference-hooks-endpoint