Claude Inference Hooks Integration

The Claude Inference Hooks integration enforces AI Security policies in real time for Claude Enterprise traffic. When the integration is active, Anthropic sends each governed prompt and tool result to your Workforce AI Security account and waits for a policy verdict before Claude processes the request. AI Security evaluates the content against your Data Loss Prevention and Attack Protection rulebase and returns an allow or deny verdict. A denied request never reaches Claude, and the user sees a blocked-by-policy message.

Because the check runs inside Anthropic's infrastructure, enforcement covers claude.ai, Claude Code, and Claude Cowork sessions across web, desktop, and CLI environments, without requiring additional proxies or agents.

You can use this integration together with the Claude Compliance API integration: the Compliance API discovers users and audits activity after it occurs, while Inference Hooks prevents policy violations in real time.

Prerequisites

  • A Claude Enterprise plan. Inference Hooks are not available on other Claude plans, on the Claude API, or on Claude models served through Amazon Bedrock or Google Cloud.

  • A Claude role that includes organization management permission (the built-in Admin, Owner, or Primary Owner roles).

  • Admin access to the AI Security portal.

How it works

  1. A user in your Claude Enterprise organization submits a prompt, or Claude receives a tool result.

  2. Anthropic sends the conversation content to your tenant's hook endpoint. Each request is signed with your organization's signing secret.

  3. Workforce AI Security evaluates the content against your rulebase and returns a verdict within the timeout configured in Claude.

  4. If the verdict is Allow, the request proceeds. On Deny, Claude blocks the request, shows the user the deny reason, and records the denial. The event also appears in the Workforce AI Security Events page with the matched rule.

Configuring the Integration

Setup requires information to be exchanged in both directions:

  • You provide Claude with your tenant's hook endpoint URL.

  • Claude provides a signing secret.

Step 1: Enable the integration in AI Security

  1. In the Check Point Portal, go to Workforce AI Security > Integrations > AI Tools & Assistants.

  2. Click Claude Inference Hooks.

  3. On the side panel, click Inference Hooks Enforcement.

  4. Copy the Hook Endpoint URL. This URL is unique to your account.

Step 2: Configure the endpoint in Claude

  1. Sign in to claude.ai as a Primary Owner.

  2. Go to Organization settings > Data and privacy > Inference hooks.

  3. Turn on Allow for your organization. This setting makes the Inference Hooks settings page available but does not enable inspection.

  4. Open the Inference Hooks settings page and click Configure.

  5. Paste the Hook Endpoint URL you copied from Workforce AI Security into the Endpoint URL field.

  6. Click Test connection to confirm Claude can reach your account.

  7. Save the configuration. Claude generates a signing secret and displays it only once. Make sure to copy it before closing the window because Claude does not display it again.

Step 3: Save the signing secret in AI Security

  1. Return to the Claude Inference Hooks panel in the Workforce AI Security portal.

  2. Paste the signing secret into the Claude Signing Secret field and click Save.

The secret is stored encrypted and is not shown again. Workforce AI Security uses the secret to verify that each verdict request comes from Claude. Requests that fail signature validation are rejected.

Step 4: Turn on enforcement in Claude

  1. On the Claude Inference hooks settings page, set Failure handling and the verdict timeout. Failure handling determines what happens if Workforce AI Security cannot return a verdict.

    • Block the request (fail closed): Claude blocks the request when a verdict is unavailable. This option provides the strongest security protection.

    • Allow the request (fail open): Claude allows the request to continue when a verdict is unavailable. This option prioritizes availability over security.

  2. Enable Enforce verdicts.

You can also use Shadow mode, which sends live traffic to Workforce AI Security and logs verdicts without blocking anyone. Review the resulting events in AI Security, adjust your rulebase as needed, and then enable enforcement. Claude also supports a rollout percentage and role-based exclusions for gradual deployment.

Limitations

  • Verdicts are allow or deny. Inference Hooks does not support prompt redaction or rewriting.

  • Attachments arrive as metadata and extracted text only. Image-only content, such as a screenshot of a document, is not inspected.

  • Anthropic currently sends only prompt-side events. Anthropic plans to add enforcement for model responses in a future release.

  • Voice mode conversations are not covered.

  • Claude system prompts and tool definitions are never included in the inspected content.

  • If Claude detects repeated failures when communicating with Workforce AI Security, it may temporarily stop requesting verdicts. During this period, policy enforcement is suspended and Claude follows the configured Failure handling behavior (fail open or fail closed) until communication with Workforce AI Security is restored.