What evidence auditors need for AI agent access (SOC 2 and ISO 27001)
How SOC 2 and ISO 27001 auditors test AI agents: the controls they map them to, the populations and samples they ask for, the gaps they find, and an evidence pack to prepare.
· 10 min read
Auditors do not have a separate rulebook for AI agents yet. In a SOC 2 or ISO 27001 audit, an agent is tested like any other non-human identity that can change things: who it is, what it may do, who approved that, what it actually did, and who noticed. The evidence they ask for is an inventory of agents and their credentials, records of access being granted and removed, approvals for high-impact actions, logs that cannot be changed unnoticed, and a change history for the rules.
What makes agents harder is volume and speed. A person makes a handful of privileged changes a week; an agent can make hundreds a day. So the question shifts from “was this one change approved?” to “can you prove every action of this kind was, and that nothing got around the check?”
The controls agents are mapped to
The SOC 2 common criteria and the ISO/IEC 27001:2022 Annex A controls below are the ones agent access usually lands on. Your auditor may map differently, so treat this as a starting point for the conversation, not the answer to it.
| Option | SOC 2 | ISO 27001 | They ask for | Evidence |
|---|---|---|---|---|
| Who the agents are | SOC 2CC6.1, CC6.2 | ISO 270015.15, 5.16, 5.18 | They ask forA list of every agent, the credentials it holds, what each may do, and who owns it. | EvidenceAn inventory of agent keys and service accounts with owner, scope, created and revoked dates. |
| Granting and removing access | SOC 2CC6.2, CC6.3 | ISO 270015.18, 8.2 | They ask forThat each agent's access was approved before it was granted, and is reviewed and removed when no longer needed. | EvidenceThe change log of keys created, rescoped and revoked, by whom, and a dated access review. |
| Authorising what agents do | SOC 2CC6.3, CC8.1 | ISO 270015.15, 8.2 | They ask forThat high-impact actions need someone's authorisation, and that the agent cannot approve its own. | EvidenceThe approval records for a sample of actions, each with who decided and when, and the rules that decided the rest. |
| Logging and monitoring | SOC 2CC7.2 | ISO 270018.15, 8.16 | They ask forThat agent activity is logged, the logs are protected, and someone looks at them. | EvidenceThe tool-call log, proof it cannot be changed unnoticed, and the alerts or reviews built on it. |
| Changing the rules | SOC 2CC8.1 | ISO 270018.32 | They ask forThat changes to what agents may do go through change management. | EvidenceWho wrote, changed or removed each approval rule and permission, with dates, matching your change tickets. |
How an auditor tests it
A Type 2 audit checks that a control worked over a period, not just that it exists. For agent actions that usually goes like this:
- You give them the population: every high-impact action the agents took in the period, such as all refunds or all production deploys.
- They test that it is complete. They compare it with a second source, such as the payment provider’s list of refunds. An action in their list and missing from yours is an exception.
- They pick a sample, often 25 to 60 items depending on how many there are, and ask for the authorisation behind each: who approved it, when, and whether it happened before the action.
- They check the rules themselves. If some actions were allowed by a standing rule, they want to see who wrote that rule, when it was approved, and that it was in force at the time.
The completeness step is where agent setups fail most often. If the approval check lives in a prompt, or only some code paths call it, the population cannot be proven complete. The check has to sit in the code that performs the action, and the tool-call log should show that no gated call ran without an approval.
Gaps auditors find
- Shared credentials
- Several agents, or an agent and a person, use the same API key. Nobody can say which one did what.
- An agent on a person's login
- The agent acts with someone's own token. Its actions look like theirs, and leave with them when they do.
- Approvals in chat
- A thumbs-up in a Slack thread is hard to find, easy to edit or delete, and not tied to the action it approved.
- Self-approval
- The process that asks can also answer, so the approval proves nothing.
- No history for the rules
- Auto-approve thresholds changed in a config file or a dashboard with no record of who, when or why.
- Logs on the same machine
- The agent's own process can rewrite or delete its logs.
- No access review
- Agent keys made for a pilot a year ago still work.
An evidence pack to prepare
Before the fieldwork starts, put these in one folder per agent, or one per kind of agent:
- What the agent is for, its owner, and the tools it can call, with the ones that need a person’s approval marked.
- Its credentials: when each was created and by whom, what it may do, and when it was rotated or revoked.
- The written approval policy: which actions run on their own, up to which amounts, which need one person and which need two. The approval policy generator writes one from a checklist.
- The change history of that policy’s rules, matched to your change tickets.
- An export of every approval in the period, and the tool-call log for the same period, so a sample can be traced from one to the other.
- The last access review, with what was removed, and the incident procedure for stopping an agent.
Where VariantGate fits
VariantGate holds the approval side of this. Each agent asks with its own key, and keys can be limited to some actions, paused and revoked; an agent key can never decide a request or touch the rules. Every decision names a person, including the ones a rule made, which are recorded in the name of the owner who wrote the rule. Keys, rules, members and the emergency stop are kept in a change log next to the decisions, and both are sealed in one hash chain anyone in the workspace can check. The record can be downloaded as CSV for the auditor on Business and above.
This is evidence for your audit, not an audit of its own. VariantGate does not make you compliant, and it does not see what your agent did after it was allowed; the agent’s report of done or failed is its own word. Your tool-call log and the system that performed the action are the other half.
Start keeping the record before the audit period does. A record cannot be filled in afterwards, and the sample will come from the whole period. More on what each entry should hold is in what an audit trail for AI agents should record.