On November 5, 2026, Anthropic shipped a Claude Enterprise security package centered on generally available inference hooks — server-side checks that send prompts and tool calls to an organization’s own DLP/security tools before Claude continues. A Security Center (beta), centralized admin console at admin.claude.com (beta), stronger access controls, and EDR sensors inside the local Cowork/Claude VM round out the release.

Coffee summary

  • FACT (Anthropic): Prompt hooks and pre-tool-call hooks are generally available for Claude Enterprise.
  • FACT: Prompt hooks forward prompts for allow/deny before generation; pre-tool-call hooks intercept tool calls (including skills/plugins/connected tools) before they run.
  • FACT: Security Center (beta) compares chat, Cowork, and Claude Code settings to Anthropic’s recommended configuration and can export an audit-style report.
  • FACT: Centralized management (beta) at admin.claude.com covers SSO, SCIM, domain verification, groups, and Compliance API keys across child orgs.
  • FACT: EDR in the local Cowork/unified Claude VM is detection-mode only (reports, does not block) and does not cover cloud Cowork sessions.

What happened

Inference hooks are the GA headline. Anthropic’s pitch is that security teams keep existing DLP policies and apply them to Claude without rewriting them as chat-only rules. Hooks run in Claude’s infrastructure rather than only on each laptop, so one org config can cover supported Claude products.

Around that core, Anthropic added admin convenience and shadow-IT controls: Admin API member sign-out (ends sessions; does not revoke API keys or kill already-running cloud sessions), self-serve IP allowlists, Desktop settings to block personal/outside Claude accounts on managed computers, and attestations inside the Security Center for shared security reviews. Anthropic stresses the Security Center is guidance, not a compliance certification.

Why it matters

Enterprise AI rollouts often stall on a simple objection: “our DLP does not see the prompt.” Moving the checkpoint server-side converts that into a vendor-integration checklist. CISOs still need to validate which products and tool paths are hooked, how fail-open vs fail-closed behaves, and what happens when the security endpoint is slow or down — details that belong in your own pilot, not in a marketing summary.

Who should care

  • Security and IT admins already running Claude Enterprise or evaluating it against ChatGPT Enterprise / Gemini for Workspace.
  • Compliance owners who need exportable configuration evidence for auditors.
  • Platform teams wiring Claude Code and Cowork where tool calls can touch production systems.

Limitations

Security Center, centralized management, Admin API sign-out, self-serve IP allowlists, Desktop lockdown, and EDR are in beta per Anthropic. EDR detects inside the local VM only. Parent-org setup at admin.claude.com cannot be undone after final approval — identify parent admins before starting. Vendor webhook lists and exact supported products live in Help Center docs that can change.

What to do next

  1. If you are on Claude Enterprise, enable inference hooks in a non-prod org first and log allow/deny decisions.
  2. Walk the Security Center page and export a report before changing production defaults.
  3. Decide fail behavior with your DLP vendor before GA traffic hits hooks.
  4. Treat Desktop personal-account blocks as a managed-device control, not a substitute for SSO.

AIImpish Take

Inference hooks GA is the substantive change; the Security Center is the readable dashboard around it. Useful for enterprises that already trust their DLP stack — not a reason to skip a scoped pilot.

Related on AIImpish

Sources: Anthropic, Claude Enterprise security updates; AI Weekly digest; secondary enterprise summaries cross-checked against the Anthropic post.