Policy Enforcement for Claude (Early Availability)
With Claude inference hooks, Anthropic sends each prompt to a Tenable endpoint. Tenable One AI Exposure scans the prompt, and the endpoint answers allow or deny. A denied prompt reaches the Claude platform, but never reaches the model, and the user who submitted the prompt sees a block message in Claude.
Blocked prompts appear in Tenable One AI Exposure as activity detections, so you can see what Tenable One AI Exposure stopped and why. Hooks work alongside the existing Claude integration monitoring, which continues to report on activity after it occurs.
The configuration steps move between Tenable One AI Exposure and the Claude Admin Console:
-
In Tenable One AI Exposure, add the Claude integration, enable Enforcement, and copy the Endpoint URL.
-
In the Claude Admin Console, add the endpoint as an inference hook, configure how it behaves, and copy the signing secret that Claude provides.
-
In Tenable One AI Exposure, paste the Signing Secret into the integration and save. Enforcement does not work until you complete this step.
-
In Tenable One AI Exposure, create the enforcement rules that determine which prompts Tenable One AI Exposure blocks.
Prerequisites
Before you begin, ensure you have:
-
A Claude Enterprise plan with administrator access.
-
The Tenable One Foundation or the Tenable One Advanced license with the Tenable One AI Governance add-on.
-
Permission to read enforcement rules to view the Enforcement Rules page.
Set Up the Claude Integration With Hooks
The Claude integration includes two features. You can turn on either individually or enable both together, but at least one must be enabled:
-
Monitoring: Monitors Claude activity, users, and policy violations using the Anthropic Compliance API.
-
Enforcement : Enforces policies in real time by connecting Claude inference hooks to your workspace.
Step 1: Add the Integration in Tenable One AI Exposure
-
In Tenable One AI Exposure, go to Settings > Integrations.
-
Add a Claude integration or edit an existing integration. For more information, see Connect Anthropic Claude.
-
In the Name and Organization ID boxes, type the corresponding values.
-
(Optional) If you enable Monitoring, type the API key in the Compliance API Key box.
-
Click the Enforcement toggle to policy enforcement.
-
Click Create ( or Save if editing an existing integration).
The Almost done! Configure the Inference Hook window appears.
-
Copy the Endpoint URL.
Step 2: Add the Inference Hook in the Claude Admin Console
-
In the Claude Admin Console, go to Settings > Data and privacy > Inference hooks.
-
Add the Endpoint URL you copied from Tenable One AI Exposure.
-
Configure the hook settings. For a description of all settings, see the Anthropic documentation on inference hooks configuration. The following settings determine how enforcement behaves:
Setting What to choose Endpoint The Endpoint URL you copied from Tenable One AI Exposure. Mode Start with Shadow mode. Claude calls Tenable and records the verdict, but allows every prompt through. Switch to Block the request after you configure the required rules. If Tenable cannot answer Select one of the following options:
-
Allow the request (fail open): The prompt reaches Claude uninspected.
-
Block the request (fail closed): Claude blocks the prompt.
Requests inspected Set a percentage from 0 to 100 percent of requests. Prompts outside the sample are not inspected, even in block mode. -
-
Save the hook.
-
Copy the signing secret that Claude provides.
Step 3: Add the Signing Secret in Tenable One AI Exposure
-
In Tenable One AI Exposure, open the Claude integration.
-
In the Signing Secret box, paste the signing secret you copied from Claude.
-
Click Save.
Without the signing secret, Tenable cannot verify requests from Claude, and enforcement does not work. The integration displays an error until you set the secret.
After you save, the integration details includes a Features section that displays Monitoring and Enforcement as either Enabled or Disabled.
For information about creating an enforcement rule, see Enforcement Rules (Early Availability).
For information about reviewing blocked activity, see Review Blocked Activity.