Dev.to Security πŸ” Cybersecurity πŸ‘ 0 πŸ“– 6 min read

AI in Your SOC 2 Environment: 3 Worked Examples, From Risk to Audit Evidence

AI shows up in SOC 2 environments in three ways: it supports the people who operate the system, it performs compliance controls, and it acts directly in production. Each one changes the risk differently, so each changes

AI shows up in SOC 2 environments in three ways: it supports the people who operate the system, it performs compliance controls, and it acts directly in production. Each one changes the risk differently, so each changes what a credible control looks like.

Below is one worked example of each, written from the auditor's side: the risk, the control, how an auditor tests it, and the evidence that shows it worked. These are illustrative designs, not a mandatory checklist. Your approach depends on your scope, risks and service commitments.

How to think about AI in your control environment

Ask how much the AI acts on its own. The more it acts without a person in between, the more your controls have to govern the AI itself, not just the people using it.

Pattern Who acts Who owns the outcome Where the risk concentrates
AI supports operations A person, using AI output The person Over-trust and volume: people approve more, faster
AI performs a control The AI, with a person overseeing The control owner Missing context and silent errors in the AI's judgement
AI acts in production The AI, on live systems and data The company, and its customers Authorisation, accuracy, and the AI straying from its purpose

Example 1: AI writes code, a human reviews it

This looks like nothing has changed. A person still owns each change, and a separate person still approves it before it reaches production.

But the risk has shifted. AI raises the volume of code and the size of each pull request, so reviewers are stretched. People over-trust AI output, and their experience tells them where humans make mistakes, not where AI does. AI makes confident assumptions with nothing behind them, and those are easy to wave through.

Risk Control Audit test Evidence
AI-assisted changes reach production without real review Branch protection requires approval from someone other than the pusher Sample merged changes; confirm approver differs from author and approved before merge Repo settings; PR history with authors, approvers, timestamps
Reviewers over-trust AI output Review guidance covering AI failure modes; review time scaled to change size; PR size limits Inspect guidance; check time-to-approve fits change size on a sample Review guideline; review comments; approval timing
Volume outpaces human review Tests, quality checks and security scanning as required checks before merge Confirm checks are mandatory, not advisory; sample changes CI config; check results
Insecure code or invented dependencies Dependency and vulnerability scanning, tracked to resolution Sample findings and trace remediation Scan reports; remediation tickets

An auditor will notice rubber-stamping patterns, like a 2,000-line change approved within a minute. Approval timing is a signal to investigate, not proof on its own. Relevant criteria: CC8.1 (change management) and CC7.1 (vulnerability detection).

Example 2: AI performs a user access review

An agent compares access lists against current staff to find leavers who still have access. In some setups it reports to a responsible person who acts. In others, it can remove the access itself.

This can make the control stronger than the human version. AI doesn't get tired or skim a long list. But it makes mistakes people wouldn't. If someone changed roles and that never reached a system the agent can see, the agent may judge their access appropriate. People pick up that context in the office. AI only has what's recorded.

Risk Control Audit test Evidence
Agent lacks context (e.g. an unrecorded role change) HR records are the source of truth for joiners, movers and leavers, and feed the agent Sample role changes and leavers; confirm each appears correctly in the agent's review HR change records; agent outputs
Agent makes a wrong call Owner reviews findings and a sample of "no issue" results before sign-off Inspect sign-off; reperform for a sample of accounts Review report; sign-off record
Agent with write access removes the wrong access Least-privilege permissions; every action logged and reviewed Inspect permissions; trace actions to findings Agent role config; action logs
A prompt, tool or model change silently weakens the control Agent configuration is under change management Sample config changes for approval and testing Change records for the agent

That last row is the one auditors care about most. When AI performs a control, the auditor needs comfort over the AI itself: how it's configured, who can change it, and whether its outputs are complete and accurate. Relevant criteria: CC6.2, CC6.3, CC4.1, CC3.2 and CC8.1.

Example 3: An agent changes live customer data

An agent runs part of a support desk as a data administrator. When a customer needs a record fixed, it makes the change in production data while talking to them. Fast and effective, and now part of your system.

Risk Control Audit test Evidence
The wrong people direct changes through the agent Authority matrix for who can request and approve each change type; requesters verified first Sample changes; confirm each requester was authorised Authority matrix; change logs with requester and approval
The wrong people can use or configure the agent Access restricted and isolated per customer environment; config access reviewed Inspect access lists and isolation Access lists; env config; review sign-off
The agent misunderstands a request Logging with before/after values; human approval for high-impact changes; reversal path Trace sampled changes to request, approval and log Tamper-resistant logs; approvals; reversals
Prompt injection or drift from purpose Scope limited to approved tools, data and change types; input guardrails; alerts on out-of-scope attempts Inspect scope and guardrails; review alert handling Tool/permission config; guardrail rules; alert records
Changes to highly sensitive data or unapproved accounts Permissions follow data classification; customers opt in first Inspect permission mapping; sample opt-ins Classification policy; permission matrix; approvals

The agent belongs in your system description too, with its purpose, boundaries and controls. Relevant criteria: CC6.1, CC6.3, CC7.2–CC7.4, CC8.1, plus Processing Integrity, Confidentiality and Privacy where in scope.

Your model provider is now a vendor

When an agent works in production, the model provider behind it is a third party to you and a fourth party to your customers. Bring it into the controls you already run for critical vendors:

  • Due diligence: review the provider's SOC 2 report and map the controls it expects you to run.
  • Data terms: confirm how the provider uses, retains and trains on your data, and where it's processed.
  • Privacy: cover it in DPAs and subprocessor disclosures where it processes personal data.
  • Change monitoring: track model and version changes, since they can change how your agent behaves.
  • Disclosure: describe the AI components in your system description.

Eight questions your auditor and customers will ask

  1. Where is AI used across development, operations, controls and product?
  2. Who owns the outcome of each AI-performed task, and how do they oversee it?
  3. What can each agent access and change, and who approved that scope?
  4. How are changes to prompts, tools and models controlled and tested?
  5. Where do agents get their context, and how do you know it's complete and current?
  6. What's logged, and could you reconstruct and reverse an agent's action?
  7. How do you detect and respond to an agent acting outside its purpose?
  8. Which model providers are involved, and what do contracts say about data use, retention and location?

The answers should point to real owners, settings and records, not just a policy saying AI is governed.

AI can do more and more of the work, but people must own the outcomes. A control is only as good as the context, scope and oversight around the AI performing it.

Originally published on the Throughline blog. Throughline is a registered CPA firm that audits AI-native companies for SOC 2, ISO 27001 and ISO 42001.

πŸ“° Read the original article on Dev.to Security

Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes β€” full credit and traffic to the original publisher.