Where AI products trip up SOC 2
| AI component | SOC 2 concern | Control approach |
|---|---|---|
| Third-party LLM APIs | Subservice organization, data retention, training on customer data | Vendor risk review, SOC 2 report review, zero-retention / no-training terms, carve-out in system description |
| Prompts & system instructions | Unreviewed changes alter system behaviour | Prompts in version control, peer review, change tickets, deployment approvals |
| RAG / vector stores | Cross-tenant data exposure | Tenant isolation, access-filtered retrieval, encryption, logging |
| Model inputs/outputs | Prompt injection, sensitive data leakage | Input validation, output filtering, DLP, red-team / pen testing |
| AI agents & tool use | Over-privileged actions | Least-privilege tool permissions, human approval for high-impact actions, audit logs |
| Fine-tuning & training data | Confidentiality and privacy of customer data | Data classification, consent, retention and deletion procedures |
Writing the system description for an AI product
Your SOC 2 system description must explain how AI components process customer data. We draft it so it is accurate for auditors and reassuring for buyers — describing model providers, data flows, retention, and the complementary controls you expect from customers and subservice organizations.
Going beyond SOC 2
SOC 2 answers the security question. Buyers increasingly also want AI-specific assurance. We layer AI governance aligned to NIST AI RMF, and ISO 42001 when certification is needed, onto the same control set.
Frequently asked questions
Does SOC 2 cover AI systems?
SOC 2 has no AI-specific criteria, but your AI features sit inside your system boundary, so the Trust Services Criteria apply to them. Auditors increasingly ask how AI components, model providers and training data are governed under existing controls.
Are OpenAI, Anthropic or other model providers in scope for our SOC 2?
They are typically subservice organizations or vendors. You need vendor risk assessments, a review of their own SOC 2 reports, data-processing terms and a clear description of how they fit into your system — usually using the carve-out method.
What AI risks do auditors and buyers ask about?
Common questions cover customer data used for training, data retention by model providers, prompt injection, output data leakage, access to model and prompt configurations, change management for prompts and models, and monitoring of AI behaviour.