Claude inference hooks are GA: wire DLP first

Anthropic published new Claude Enterprise security and admin controls on November 5, 2026. The change that is ready to use in production is inference hooks: prompt hooks and pre-tool-call hooks are generally available, so prompts and tool calls from supported Claude products can be checked by the data security tools your organization already runs. The decision for security and platform owners is whether to connect those hooks to existing DLP before they expand Claude access further. Anthropic limits both hook types to "supported Claude products" and does not name which products those are.

The catch is packaging. The same announcement puts the Security Center, centralized management, the self-serve IP allowlist, Admin API member sign-out, and the Endpoint Detection and Response (EDR) integration in beta. Anthropic's article is explicit that the Security Center is guidance against Anthropic's recommended configuration, "not a certification or a guarantee of compliance with any law or standard."

What are Claude inference hooks?

Inference hooks connect Claude to your existing security stack so data loss prevention (DLP) and related policies can apply as people work. Anthropic's Enterprise security updates article describes two kinds, both generally available:

• Prompt hooks send prompts from supported Claude products to your security tool for a check.
• Pre-tool-call hooks check tool calls in supported Claude products before they run.

The article's pitch is practical: "Your team keeps the policies it has already written and applies them to Claude," and it says hooks work with security vendors.

BaristaLabs interpretation: hooks matter most on supported Claude products where people paste customer data into prompts or where tool calls can leave the machine. A GA check on prompts and tool calls is a stronger production control than a beta dashboard that compares your settings with recommendations.

What else shipped on November 5?

The Help Center release notes date the Enterprise security and admin package to November 5, 2026, and point to the same article. Beside hooks, that article covers:

• Security Center (beta): one Organization settings page that reviews security settings for chat, Cowork, and Claude Code against Anthropic's recommended configuration, with current values, recommended values where available, links to change each setting, attestations for shared review, and a downloadable report. Organizations with the Compliance API can read the same results through that API.
• Centralized management (beta): a parent-organization console at admin.claude.com for company-wide SSO, SCIM, domain verification, groups, and Compliance API keys across Claude Enterprise and Claude Platform organizations. Setup "can't be undone," so Anthropic tells admins to identify parent organization admins before they start.
• Access controls (beta): Admin API member sign-out and self-serve IP allowlisting are in beta. Owners can also manage Claude Desktop settings in Organization settings (beta), including blocking personal and other outside Claude accounts on managed computers.
• EDR in the local Cowork / unified Claude VM (beta): your organization's sensor can run inside the local virtual machine in detection mode. It "report[s] activity to your security team and [doesn't] block it," and it covers the local VM, not Cowork sessions in the cloud.

How is this different from Claude Security scanning?

Claude Security is Anthropic's repository vulnerability scanning product, which BaristaLabs covered in an earlier post. Inference hooks and the Security Center are a different layer: they govern how people use Claude products day to day, and whether prompts and tool calls meet your DLP rules, not whether a repository has a security finding.

BaristaLabs interpretation: treat them as complementary. Scanning finds issues in code. Hooks check prompts and tool calls against your policies while someone works. The Security Center helps you see which product settings still sit away from Anthropic's recommendations.

Who should enable hooks now?

BaristaLabs recommendation: enable inference hooks first if you already run Claude Enterprise and you already have a DLP or security vendor whose policies you want applied to Claude. That is especially true if you are about to expand Claude access on surfaces that handle customer data or tool calls. Confirm with Anthropic or your vendor which of your surfaces count as supported Claude products for hooks. Our earlier post on Claude Code's cloud and Remote Control role split covers the access side of the same decision.

Wait, or pilot narrowly, if any of these apply:

• No DLP owner for Claude yet. Hooks only help if a tool receives the check and someone acts on the result. Assign an owner before you flip the switch.
• You expected EDR to block unsafe tool use. Anthropic's article says sensors are detection-only inside the local VM.

IP allowlisting is a separate decision from hooks, so plan it separately. The Enterprise self-serve IP allowlist in this announcement is still beta, and yesterday's Console IP allowlist walkthrough covers a separate Console path. Hooks are not a substitute for either.

Recommended first step: one hook surface, then Security Center

BaristaLabs recommendation: ship one production control, then use the beta dashboard as a checklist.

1. Inventory Claude surfaces in use: chat, Claude Code (local, cloud, Remote Control), Cowork, and desktop. Note which ones handle customer data or tool calls, then confirm which of those Anthropic or your vendor treat as supported Claude products for hooks.
2. Connect prompt hooks and pre-tool-call hooks to the security tool that already owns DLP for email or SaaS, starting with the highest-risk surface among the supported Claude products.
3. Pilot with a small group for a week. Confirm that legitimate work still completes and that hook checks on prompts and tool calls show up where your security team looks.
4. Open the Security Center once after hooks are live. Use it to list gaps against Anthropic's recommended configuration, export the report if auditors need it, and track attestations. Do not treat a clean report as a compliance certificate; Anthropic says it is not one.
5. Hold centralized management setup until you have named the parent organization admins, because Anthropic says that setup cannot be undone.

For help planning AI tool controls with your existing security stack, see AI consulting.

Sources and scope

• More visibility and control over Claude Enterprise for security teams, Anthropic / Claude, dated with the November 5, 2026 Help Center entry
• Release notes, Claude Help Center, November 5, 2026 entry

BaristaLabs did not enable inference hooks or open a Security Center on a test Enterprise organization for this post. Availability, vendor integrations, and beta status come from Anthropic's article and Help Center as of November 6, 2026. Lines marked BaristaLabs interpretation or recommendation are BaristaLabs' views, not the vendor's.