>
    Strata Copilot
    Integrate Anthropic Inference Hooks
    Focus
    Download PDF
    Focus
    1. Home
    2. Prisma AIRS
    3. AI Runtime Security
    4. AIRS VM Firewall
    5. Integrate Anthropic Inference Hooks
    Download PDF
    Prisma AIRS

    Integrate Anthropic Inference Hooks

    Table of Contents

    Integrate Anthropic Inference Hooks

    Learn about Prisma AIRS integration with Anthropic Inference Hooks
    Where Can I Use This?What Do I Need?
    • Prisma AIRS- AI Runtime Security in AWS
    • Prisma AIRS Licenses
    Prisma AI Runtime Security (AIRS) integrates with Anthropic Inference Hooks to inspect and govern every Claude prompt before it reaches the model. When a governed prompt arrives, Anthropic pauses inference, POSTs the conversation transcript to a Prisma AIRS endpoint, and enforces the allow or deny verdict AIRS returns — all before the model ever reads the message.
    This feature reverses the normal call direction: Anthropic calls Prisma AIRS.
    This topic covers how the integration works, how to configure it, what data flows between the systems, and operational considerations.
    This feature is limited to the US region only.
    Prerequisites
    Before configuring the integration:
    1. The AIRS webhook endpoint must be publicly reachable on HTTPS port 443 with a publicly-trusted certificate. Anthropic does not follow redirects.
    2. If your environment restricts inbound sources, allowlist 160.79.106.0/25 — the CIDR range Anthropic uses to send webhook requests.
      IP allowlisting narrows exposure but is not a substitute for signature verification.
    3. Ensure that a configured AIRS API key is linked to a configured AIRS scan profile.
    Lightweight profiles (prompt injection, topic guardrails) are recommended for latency-sensitive gating; heavy profiles (DLP, malicious-code detection) add latency.

    Setup and Configuration

    1. Log in to Strata Cloud Manager and navigate to AI SecurityAI RuntimeAPI Applications.
      You can configure inference hooks using options for API Applications, AI Sessions, or API Violations.
    2. Click Manage in the upper-right of the AI Applications page.
    3. Select Applications.
    4. Click Add New Application/API Key.
      Use Edit to configure an existing deployment profile.
    5. In the Onboard API Account dialog, select the profile to begin API onboarding and click Next.
    6. In the Add Application (Agent) screen, specify the following details and click Next.
      1. Enter an Application Name.
      2. Select a Cloud Provider from the drop-down.
      3. Select the Environment.
      4. Select the AI Agent Framework.
      5. Enter a Webhook Secret.
      6. Enable the Security Profile toggle and use the drop-down to link an existing security profile or choose which profile to use in your app code.
    7. In the Input API Details page, configure the API key and click Generate API Key.
      1. Enter an API Key Name.
      2. Select the key Rotation interval from the drop-down.
      3. Click Generate API Key.
        The Implement API Key screen appears with the generated API key:
    8. Click Done.
      Next, configure Anthropic Inference Hooks in the Anthropic console.
    9. In the Anthropic console, go to Organization settingsInference hooks and configure the following:
      1. Set the endpoint URL to the AIRS webhook URL provided in the Prisma AIRS console: https://service.api.aisecurity.paloaltonetworks.com/inference-hooks/scan/sync/request
      2. Add a custom header carrying your AIRS API key:
        • Header name: x-pan-token
        • Header value: your AIRS API key
        Anthropic reserves webhook-*, X-Anthropic-*, and Content-* prefixes, and requires hyphens (not underscores) in header names. x-pan-token complies with these rules.
    10. Return to the Prisma AIRS console and update the application configured in the previous steps to enter the signing secret displayed by Anthropic after saving the Inference Hooks configuration.
      The signing secret is shown only once. During approximately the first minute after a secret is set or rotated, Anthropic may send requests that are unsigned or signed with the previous secret. Prisma AIRS tolerates this propagation window to avoid blocking prompts across the organization during activation.
    11. Test and activate the integration:
      1. In the Anthropic console, click Test connection. Anthropic sends a synthetic config-test frame. Prisma AIRS returns allow for these frames without logging.
      2. Once the test succeeds, enable Enforce verdicts in the Anthropic console. Enforcement propagates within approximately one minute.

    Runtime Flow

    For each governed prompt, Anthropic:
    1. Pauses inference.
    2. POSTs a signed PromptFrame to the AIRS webhook endpoint.
    3. Waits for a Verdict response within the configured verdict timeout (1–10,000 ms).
    4. Enforces the verdict: allow continues inference; deny blocks it and shows the deny_reason to the user.
    Known Limitations
    • Text content only. Images, files, and binary attachments are not scanned. tool_use and tool_result blocks are carried but not inspected.
    • Prompt event only. Response-side enforcement is not yet supported.
    • US region only.
    • surface.application is advisory. This field is client-asserted and cannot be used as a security control.

    © 2026 Palo Alto Networks, Inc. All rights reserved.