By Hamza Faisal

Anthropic has handed enterprise customers a veto over Claude. In a security update published on November 5, the company made “inference hooks” generally available on Claude Enterprise, letting organisations send prompts and tool calls to their own data-loss prevention tools for a check before Claude generates a reply or runs a tool. If the customer's security tool says no, Claude does not answer.

The move targets the objection that has kept the most cautious companies — banks, hospitals, law firms — from letting staff use AI assistants on real work: once a prompt leaves the building, nobody on the security team can see it, let alone stop it. Anthropic's answer is to put the customer's own security stack in the request path.

Two checks before Claude acts

There are two kinds of hook. Prompt hooks send a user's prompt to the organisation's security tool before Claude generates a response. Pre-tool-call hooks inspect a proposed tool call — reading a file or querying a database — before it executes. Verdicts are binary — allow or block — with no rewriting or partial redaction, and the default timeout is five seconds, adjustable by the customer.

The payload covers conversation text, tool calls, their results and attachment text. File and image bytes and system instructions are excluded. Anthropic says it keeps only configuration data and verdicts, not the inspected content. The hooks work with data-loss prevention endpoints from Netskope, Palo Alto Networks, Proofpoint and Zscaler, or with a customer's own internal server — and they apply across claude.ai on web, desktop and mobile, plus Cowork, Claude Code, Claude for Microsoft 365, Claude in Slack and Claude in Chrome.

A security centre and one console for the whole company

The hooks arrive alongside a batch of controls still in preview. A new security centre, in beta, lines up an organisation's Claude settings for chat, Cowork and Claude Code against Anthropic's recommended configuration, flags the gaps and exports an auditor report. Centralised management, also in beta, lets a parent organisation run single sign-on, directory sync, domain verification, groups and Compliance API keys for its Enterprise and Platform organisations from one console at admin.claude.com.

Endpoint monitoring is in preview too: sensors from CrowdStrike, SentinelOne, Palo Alto Networks and Tanium can now run inside Cowork's local virtual machine on macOS and Windows. They detect only — they do not block Cowork tasks — and they do not cover cloud sessions. Anthropic is charging no extra fee for the preview. New access controls let administrators sign a member out of all devices with one Admin API call, run an IP allowlist, and block personal Claude accounts in Claude Desktop on managed computers.

Why regulated firms were waiting for this

The timing is deliberate. Anthropic is pitching Claude to the largest, most regulated companies in the world as it prepares for a public listing — and for those buyers, the question is never whether the model is clever enough, but whether the security team can prove nothing sensitive leaked. Putting a customer's own data-loss prevention tools in the loop answers that question in the customer's own language, reusing policies already written rather than recreating them inside Claude.

It also sharpens the enterprise race. OpenAI, Google and Microsoft all sell admin and compliance controls of their own, but Anthropic's pitch is that the check happens before generation, not after the fact in an audit log. Whether that wins over the regulated firms that have held out — and how fast the preview controls graduate to general availability — will be the test in the months ahead.