Anthropic lets enterprises veto Claude responses before they are generated

Summary

• Anthropic opened Claude Enterprise prompts and tool calls to customers' own data-loss prevention servers before execution.
• Inference hooks are generally available; a security hub, central management and EDR monitoring remain in preview.
• The change removes a core blocker for regulated firms that could not police AI output in real time.

The latest

Enterprise customers can now route every Claude prompt through their own security server and block the response before it is generated, under controls Anthropic announced on November 5, 2026. The feature, called inference hooks, is generally available on Claude Enterprise plans and is joined by a security hub, centralized administration and new access controls, most of them still in preview.

Details

• How it works: Anthropic said inference hooks send a request to a customer-selected security server before Claude generates a reply, and Claude proceeds only on an allow verdict. A second event inspects tool calls before they execute. Verdicts are binary — allow or block — with no rewriting and no partial redaction.
• Failure handling: The default timeout is 5 seconds and is adjustable, according to Anthropic. If the customer's security server goes down, the administrator decides in advance whether traffic is blocked outright or allowed through uninspected — a choice that trades availability against exposure.
• What is sent: The payload covers conversation text, tool calls, their results and attachment text, Anthropic said. It excludes file and image bytes and system instructions. Anthropic said it retains only configuration data and verdicts, not the inspected content itself.
• Named vendors: Supported data-loss prevention endpoints include Netskope, Palo Alto Networks, Proofpoint and Zscaler, or a customer's own internal server — aiming the feature at firms already running these tools across email and cloud traffic.
• Coverage and gaps: Hooks apply to claude.ai on web, desktop and mobile, plus Cowork, Claude Code in the terminal, Claude for Microsoft 365, Claude in Slack and Claude in Chrome. Image-only content is out of scope, as are Bedrock users, Google Cloud users and Claude Platform API organizations.
• Security hub: In preview, the hub compares an organization's settings across chat, Cowork and Claude Code against Anthropic's recommended configuration, links to each setting and exports an auditor report. Anthropic calls it "guidance, not a compliance certification" and guarantees no legal or standards compliance. The full control list is unpublished.
• Central management: A parent organization can run single sign-on, directory sync, domain verification, groups and Compliance API keys through admin.claude.com for both Claude Enterprise and Claude Platform organizations. Child organizations keep their own members and controls. Setup requires every affected administrator's approval and cannot be reversed.
• Endpoint monitoring: Preview support on Team and Enterprise lets EDR sensors run inside Cowork's local virtual machine on macOS and Windows, with CrowdStrike, Evoke Security, Neo, Palo Alto Networks, Pluto Security, SentinelOne and Tanium approved. Claude Desktop verifies the vendor package signature at each VM launch.
• Detection only: Those sensors detect but do not block Cowork tasks, and do not cover cloud Cowork sessions, Anthropic said. No additional Anthropic fee applies during the preview period.
• Access controls: Administrators can sign a member out of all devices with one Admin API call without removing them, though it does not revoke API keys, running cloud sessions or scheduled tasks. Owners can also restrict Claude Desktop to approved organizations, block personal accounts, disable auto-updates, limit file tools to named folders and bar user-added local MCP servers.

Between the lines

The binary allow-or-block design and the exclusion of file bytes, images and system instructions mark the limits of what the inspection layer sees. Pairing generally available hooks with a preview security hub Anthropic itself calls guidance rather than certification suggests the enforcement piece is ready for production while the audit layer is not.

What's next

Anthropic has not announced when the preview features — the security hub, central management, Cowork EDR sensors and the new access controls — move to general availability, or whether the EDR integration will carry a fee once the preview ends.

Source: Anthropic