Anthropic lets Claude Enterprise customers route prompts through their own data-loss tools

What Anthropic released

On 5 November 2026, Anthropic announced new security features for business customers. As BornCity reported, the one that is ready for general use is called inference hooks, available in Claude Enterprise. Inference is simply the model producing an answer, and a hook is a point where a company's own software can step in.

In this case, the hooks let an organization pass what employees type, and what Claude plans to do with connected tools, to the company's existing data loss prevention systems. Those are tools built to stop confidential information from leaking. The point is that security rules a company already has can keep working alongside Claude without a rebuild.

How the hooks work

There are two kinds. Prompt hooks catch what a user types. Pre-tool-call hooks catch an action Claude intends to take with a tool, such as reaching into another service. In both cases the content is intercepted before it runs and can be checked by a third-party tool.

A single hook configuration covers many places where people use Claude: claude.ai, Cowork, Claude Code, Microsoft 365 integrations, Claude Tag and Chrome extensions. That matters because security teams usually dislike setting up separate rules for every product.

Each check has a time limit. The default is 5 s, or 5,000 ms, and it can be set anywhere between 1 and 10,000 ms. The system also blocks public tunneling services such as ngrok, which expose a private computer to the internet, for security reasons.

Who can use it, and who cannot

Access is limited. Hooks are for Claude Enterprise customers only and cannot be used through Amazon Bedrock or Google Cloud. Setting them up requires the organization:manage permission, which belongs only to primary owners and owners. Ordinary administrators are shut out.

That owner-only rule is worth noticing. In many companies the people who run day-to-day IT are not the account owners, so rolling this out may need a conversation with whoever holds that top role.

The rest of the package is still in beta

Anthropic announced other controls that are not yet generally available. The Security Center brings settings for chat, Cowork and Claude Code into one screen. It compares them with Anthropic's recommendations and lets administrators record risks they have chosen to accept. Reports, including confirmations, can be exported as a PDF.

Anthropic was explicit that the Security Center is not a formal certification or a legal guarantee of compliance. Individual checks are labeled Passed, Failed, Not attested or Shown for information, and viewing them is again limited to the owner role.

Other beta items include:

• A parent console at admin.claude.com to manage several Claude organizations together, covering single sign-on, automated user provisioning through SCIM, domain verification and groups. Setting it up is permanent and cannot be undone later.
• An IP allowlist with up to 1,000 entries, limiting access to approved network addresses.
• A remote sign-out function in the admin API that ends active sessions across browsers and devices. It does not cover API keys, running cloud sessions or scheduled routines.
• A setting to block personal or external accounts in Claude Desktop on managed computers.
• An endpoint detection and response integration (software that watches a computer for threats). It runs in a local Cowork virtual machine in detection-only mode and does not capture sessions that run purely in the cloud.

What to read into the gaps

The source describes capabilities, not results. It does not say what happens if a hook check runs out of time, which third-party tools have been tested, or what any of this costs. Those details decide whether hooks work as a real safeguard or only as a checkbox.

The beta items also carry limits that are easy to miss. A remote sign-out that skips API keys, for example, would not stop every kind of access. Security teams should read those exclusions as part of the design and plan around them.