---
title: What SOC 2 evidence do auditors want for AI agents?
url: https://opencontroller.lyzr.com/blog/soc-2-evidence-for-ai-agents
published: 2026-11-12
updated: 2026-11-12
author: Lyzr
---

# What SOC 2 evidence do auditors want for AI agents?

**SOC 2 has no AI-specific criteria, so auditors test agents with the criteria they already use: access (CC6), operations and monitoring (CC7), change management (CC8) and vendor risk (CC9). Expect them to ask who can change an agent, what identity it acts under, how you detect misuse, and how you assess the model provider.** For a Type 2 report, the evidence must cover the full observation period.

SOC 2 is an attestation by a CPA firm against the AICPA Trust Services Criteria (2017, with points of focus revised in 2022). Security is always in scope. Availability, Processing Integrity, Confidentiality and Privacy are optional.

## Which Trust Services Criteria apply to AI agents?

| Criteria | What auditors test | Agent-specific question | Sample evidence |
|---|---|---|---|
| CC1, CC2 Control environment, communication | Policies and responsibilities | Is there an agent policy and an owner per agent? | Approved policy; registry export with owners |
| CC3 Risk assessment | Risks identified and assessed | Were agent risks such as prompt injection and tool misuse assessed? | Risk register entries; threat models |
| CC6 Logical access | Access granted, reviewed and removed | What identity does the agent use, and who can administer it? | SSO configuration; agent identity mapping; access reviews |
| CC7 System operations | Monitoring, anomaly detection, incident response | How do you detect an agent misbehaving and respond? | Alert history; incident tickets; quarantine records |
| CC8 Change management | Changes authorised, tested, approved | Is a new prompt, model or tool a tracked change? | Version history; evaluation results; approvals |
| CC9 Risk mitigation | Vendor and business-partner risk | How do you assess model providers and agent platforms? | Vendor assessments; their SOC 2 reports |
| PI1 Processing integrity | Complete, accurate, timely processing | How do you know agent output is accurate? | Groundedness or quality scores |
| C1 Confidentiality | Confidential data protected | Does confidential data leave through prompts? | Redaction policy; gateway logs |

## Which agent controls fail SOC 2 testing most often?

1. **Prompt and tool changes outside change management (CC8).** A model swap or prompt edit is pushed live with no ticket. Treat agent versions like code releases.
2. **Shared API keys (CC6).** Several agents use one key, so access cannot be attributed or revoked per agent.
3. **Orphaned agents (CC6).** The builder leaves; the agent keeps running with their access. See [what happens when an agent's owner leaves](/blog/what-happens-to-ai-agents-when-the-owner-leaves).
4. **Monitoring that nobody reviews (CC7).** Logs exist, but there is no evidence of review or response.
5. **Model providers missing from vendor management (CC9).** Usage started on a corporate card and was never assessed.

## How do you prepare agent evidence for a Type 2 audit?

Use this checklist at least one quarter before the observation period starts.

1. **Define the boundary.** List which agents are part of the system in the report. Discover any you have missed.
2. **Assign owners** to every in-scope agent in the registry.
3. **Move agents to their own identities**, federated through your identity provider, with SSO for admin access.
4. **Put agent changes through change control**: version, evaluate, approve, release.
5. **Route traffic through a gateway** so access, policy decisions and cost are logged consistently.
6. **Set alert review** with a named reviewer and a response time, and keep the tickets.
7. **Add model providers and agent platforms** to the vendor inventory and collect their reports.
8. **Rehearse one incident**, such as quarantining an agent, and keep the record.

Auditors sample. Every in-scope agent should pass, not just the flagship one.

## What if your customer asks for your AI vendor's SOC 2?

Ask any AI governance or agent vendor for a current SOC 2 Type 2 report and bridge letter, and check that the system description covers the product you are buying. Our list of [security questions for AI governance vendors](/blog/security-questions-to-ask-an-ai-governance-vendor) covers the rest.

## How Opencontroller does this

Opencontroller provides evidence for SOC 2 controls; your auditor decides sufficiency. A governed registry of agents, versions, tools and MCP servers supports change management (CC8). OIDC SSO, delegated on-behalf-of identity and agent workload identity federation support access controls (CC6). Posture findings, gateway logs, LLM judges on production traffic and rogue-agent quarantine and kill support operations and monitoring (CC7). See the [framework mapping hub](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Does SOC 2 have AI-specific criteria?**
No. Agents are tested under the existing criteria, mainly CC6, CC7, CC8 and CC9.

**Are AI agents in our SOC 2 scope?**
If they are part of the system described in the report, yes, and auditors may sample them.

**How long must agent evidence cover?**
For a Type 2 report, the whole observation period, commonly between 3 and 12 months.

## Key facts

SOC 2 is an AICPA attestation report that tests controls against the 2017 Trust Services Criteria (points of focus revised in 2022). There are no AI-specific criteria; AI agents in scope are tested through the common criteria, chiefly CC6 (logical access), CC7 (system operations, monitoring and incident response), CC8 (change management) and CC9 (risk mitigation, including vendors), plus Processing Integrity and Confidentiality where selected. A Type 2 report tests operating effectiveness over an observation period, commonly 3 to 12 months, so agent evidence must cover the whole window. Opencontroller by Lyzr provides evidence such as a governed registry with versions, OIDC SSO, delegated identity, Posture findings, gateway logs and quarantine events. Pricing starts from $250K per deployment per year.

## Related

- [How do you map AI agent controls to compliance frameworks?](https://opencontroller.lyzr.com/blog/how-to-map-ai-agent-controls-to-compliance-frameworks.md)
- [What security questions should you ask an AI agent governance vendor about itself?](https://opencontroller.lyzr.com/blog/security-questions-to-ask-an-ai-governance-vendor.md)
- [What happens to AI agents when their owner leaves the company?](https://opencontroller.lyzr.com/blog/what-happens-to-ai-agents-when-the-owner-leaves.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
