Skip to main content

Pro audit program · v1.0

AI Vendor & Data-Sharing Risk

Most "AI risk" is really 3rd-party risk: data egress to OpenAI/Anthropic/Google + training opt-outs + retention. Quick check.

  • General target area
  • ISO 42001 / Vendor Risk framework
  • 7 controls in this program
  • Cyentrix Cyentrix Trusted Author
Sign in to run → Back to library

About this program

Most “AI risk” is really 3rd-party risk: data egress to OpenAI/Anthropic/Google + training opt-outs + retention. Quick check.

Risks addressed

  • Critical Confidential data exfiltrated via free-tier AI tools
  • Critical Data used to train the provider next model
  • High Output relied on without verifying provenance / accuracy

Controls (7)

  1. AI-vendor register with data flows

    High

    An AI-vendor register with data flows is a centralized inventory documenting all third-party artificial intelligence systems, services, and providers used across the organization, including detailed mappings of what data is sent to each vendor, how it is processed, and where…

    How to test + evidence

    Testing procedure: Every AI vendor in use mapped to what data is sent.

    Evidence to collect: Vendor register.

  2. Enterprise tier / DPA covering training opt-out

    Critical

    This control ensures that an organization using enterprise-tier AI services (such as GitHub Copilot Enterprise, OpenAI API, or similar large language model platforms) has negotiated and documented a Data Processing Agreement (DPA) or addendum that explicitly opts out of using…

    How to test + evidence

    Testing procedure: Confirmation in contract that customer data is NOT used to train the provider models.

    Evidence to collect: Signed DPA / contract clause.

  3. Block / proxy free-tier consumer AI tools

    High

    This control blocks or proxies access to consumer-grade, free-tier generative AI tools (e.g., ChatGPT free, Google Bard, Claude.ai) through network-layer enforcement, web filtering, or egress gateway policies. Organizations configure DNS filtering, TLS inspection proxies, or cloud access security brokers (CASB)…

    How to test + evidence

    Testing procedure: Egress controls block consumer AI domains for corporate devices, or proxy through approved gateway.

    Evidence to collect: Egress policy + DLP.

  4. Approved-tool allowlist communicated to staff

    High

    This control ensures that an organization maintains and actively communicates a documented list of approved software tools, applications, and utilities to all relevant staff members. The allowlist defines which tools employees are permitted to install and use on organizational systems,…

    How to test + evidence

    Testing procedure: Staff know what they can use and what they cannot; reminders + training.

    Evidence to collect: Policy + training material.

  5. PII / IP not sent to AI without classification check

    Critical

    This control requires that all data transmitted to artificial intelligence systems, including large language models and machine learning platforms, undergoes automated classification checks prior to transmission to identify and block Personally Identifiable Information (PII) and Intellectual Property (IP). The classification…

    How to test + evidence

    Testing procedure: DLP scans uploads / pastes to AI tools for Restricted-classified data.

    Evidence to collect: DLP policy + sample alert.

  6. Retention limit + log purge on AI vendor side

    Medium

    This control ensures that AI service providers automatically delete or anonymize logs, prompts, model inputs, outputs, and telemetry data according to contractually defined retention periods. It requires vendors to implement automated purge mechanisms that remove customer data from production systems,…

    How to test + evidence

    Testing procedure: Vendor retention configured to minimum or zero where possible.

    Evidence to collect: Vendor retention setting.

  7. Provenance + human-review of AI output for critical use

    High

    This control ensures that AI-generated outputs used in security-critical decisions or operations are tagged with provenance metadata (model identity, version, timestamp, prompt hash) and subjected to mandatory human review before deployment or action. Organizations maintain an inventory of critical AI…

    How to test + evidence

    Testing procedure: Code / legal / medical / customer comms from AI reviewed by qualified human before use.

    Evidence to collect: Review process doc.