AI Guardrails for SOC 2 and ISO 27001 Audit Evidence
TL;DR: Auditors sample records that show a control operated, so a list of configured AI guardrails is only a starting point. An AI gateway produces two kinds of control evidence. The audit log records administrative changes, such as new keys or guardrail edits, and no user can edit or delete its entries. Policy decision records capture each allow, block, modify, or redact verdict with a correlation ID, the policy version, and the caller. Map both records to the SOC 2 Trust Services Criteria or ISO 27001 Annex A controls you report against. Export them to your security information and event management (SIEM) system, with retention set at that destination. Business-logic checks inside your applications still need their own evidence.
Security teams can usually list their AI guardrails. An auditor asks a harder question: which record shows that each control operated during the review period? A policy file shows how a control was designed. A SOC 2 Type II or ISO 27001 auditor testing operating effectiveness wants records from the period itself.
This guide covers the records an AI gateway produces, the SOC 2 and ISO 27001 controls each record can support, what the gateway can’t prove, and how to prepare the records before the auditor asks.
What auditors need beyond a list of AI guardrails
Confidence in an access control is easy to report and hard to back with records. Cequence Security and Enterprise Management Associates found that 94% of the enterprise IT and security leaders they surveyed were at least moderately confident their AI agents had no more access than they needed. Only 33% said they provision agents with least-privilege access.
An audit tests exactly that gap. The auditor picks a sample of agent requests from the review period. For each one, the auditor asks for a record that shows which identity made the request, which policy applied, and what the policy decided.
Application-level guardrails make that sample hard to produce. Each application logs in its own format. A call that skips the sanctioned library leaves no record at all. A gateway that every model call and tool call passes through writes one record format for every request, which makes sampling practical.
Two records an AI gateway produces
Tetrate Agent Router Enterprise writes two separate record streams. Each stream answers a different audit question.
| Record | What it captures | Audit question | Can it be changed? |
|---|---|---|---|
| Audit log | Administrative actions: API keys, guardrail changes, membership edits | Who changed a control, and when? | No user can edit or delete an entry |
| Policy decision record | One verdict per request: allow, block, modify, or redact | Did the control run on this request? | Protection depends on the export destination |
The audit log is change-management evidence. Tetrate’s docs state that no user can edit or delete an entry, regardless of role. Signed Merkle checkpoints make tampering detectable in sealed windows of the log.
Policy decision records are operating evidence. Each record carries a correlation ID, the policy version in force, the caller identity, the model, and the action taken. A decision record exists only when at least one policy is in force, such as a guardrail that blocks or redacts, a rate limit, or an enforced budget. A deployment with no policies produces audit events but no decision records.
For how both streams sit alongside data-leakage controls, see AI data leakage prevention at the gateway.
Mapping guardrail records to SOC 2 and ISO 27001
The SOC 2 Trust Services Criteria (TSC) are the criteria a SOC 2 report tests controls against. Annex A of ISO/IEC 27001:2022 lists 93 reference controls. A gateway record can support a control, but the auditor decides what counts as evidence. Confirm this mapping during audit planning.
| Guardrail | Record it produces | SOC 2 | ISO 27001 Annex A |
|---|---|---|---|
| PII redaction | Redact or modify verdict per request | CC6.7 | A.8.11 data masking, A.8.12 data leakage prevention |
| Identity on every request | Caller identity in each decision record | CC6.1 | A.5.15 access control |
| Tool access control | Allow or block verdict per tool call | CC6.1 | A.5.15 access control |
| Guardrail and key changes | Audit log entry | CC8.1 | A.8.32 change management |
| Export to your SIEM | Both streams in your own logging system | CC7.2 | A.8.15 logging, A.8.16 monitoring activities |
Each decision record names the caller by API key and, where available, by user. If your auditor needs a named person for each request, check which of your applications send a user identity.
Healthcare teams can also use redact records as input to HIPAA audit controls, which require mechanisms that record and examine activity in systems that hold electronic protected health information.
Tool calls are where least privilege gets tested
Agents reach production systems through tools, so auditors test least privilege for AI at the tool layer. The Model Context Protocol (MCP) lets an agent call databases, ticketing systems, or internal APIs. Without a gateway in the path, each MCP server applies its own access rules, and each one keeps its own logs.
Agent Router Enterprise routes tool calls through an MCP gateway with a curated tool catalog. Each access profile authenticates separately, and each profile sees only the tools assigned to it. Security teams can then audit every tool call an agent makes, alongside every model call.
Step-up authorization, built with Ory, covers higher-risk agent actions. Routine actions run without interruption. A higher-risk action waits for human approval through your own identity system, so a person signs off before the agent acts.
The MCP audit logging guide covers what to capture for each tool call.
What the gateway can’t prove
A gateway sees requests and responses. It can’t see the business rules your application applies before or after a model call.
A check that a payment falls within a user’s approved limit runs inside the application. So does a warning shown before a user submits a sensitive prompt. The gateway has no record of either check, so the application team needs its own evidence for those controls, such as application logs or test results from the review period.
| Layer | Evidence it produces | Gap |
|---|---|---|
| Gateway | Decision records for every request it handles, plus the audit log | Business logic, and traffic that bypasses the gateway |
| Application | Logs and tests for workflow-specific checks | Only instrumented applications, with formats that vary by team |
Traffic that bypasses the gateway is a gap in both layers. Egress rules that block direct calls to model providers narrow that gap, and the egress rule is evidence in its own right.
How to get AI guardrails audit-ready
Each of the five steps produces an artifact an auditor can test. Start them before the review period begins, because a gateway can’t create records for requests that already ran.
Step 1: Inventory AI traffic before writing policy
Review network egress logs, cloud provider bills, and application dependencies to find every model, provider, and tool in use. Include unsanctioned use as well as sanctioned use.
Record which teams use which models. Note the data classifications that go to each provider, then list the agents that call tools over MCP. The inventory sets the scope of the audit, because the gateway can’t produce records for AI traffic that never passes through it.
Step 2: Map each control to a clause
Start from the mapping table above. Agree with your auditor during planning on which record samples which control.
Then name an owner for each control. You configure some controls yourself, such as guardrail rules, retention, and egress rules. Tetrate operates others, such as the security of the management plane. You verify those through Tetrate’s SOC 2 Type II report, available through the Trust Center. Keep the two lists separate, because your auditor tests the first list directly and reviews the second through the vendor report.
Step 3: Run new rules in monitor mode first
Run each new guardrail rule in monitor mode before you enforce it. Monitor mode logs violations without blocking requests, so a team can tune detection thresholds without breaking developer workflows.
Auditors often ask how a threshold was chosen. The monitor-mode results from the tuning period answer that question. The audit log records each guardrail change, so the switch from monitor mode to enforcement appears with its date.
Step 4: Export both streams and set retention at the destination
Export the audit log and policy decisions to Amazon S3, Splunk, or Datadog.
Exported copies follow the destination’s retention policy. Purging a record in Agent Router Enterprise doesn’t delete copies that were already exported. Set retention and write-once protection at the destination, for example with Amazon S3 Object Lock. ISO 27001 control A.8.15 asks that logs be protected against tampering. SOC 2 doesn’t prescribe a storage method, so agree the protection level with your auditor.
Step 5: Rehearse an auditor sample
Before the audit, pull one block record by its correlation ID. Show the policy version in force when the request ran. Then show the audit log entries for changes to that guardrail during the review period.
If your team can’t connect those records quickly, fix the export or the retention settings before fieldwork starts.
Choosing a data capture mode: evidence vs data minimization
A richer record makes stronger evidence, but it also stores more sensitive data. Agent Router Enterprise lets each customer choose whether the gateway stores the full request, only counts and cost, or nothing. You can change the setting at any time.
| Mode | What reaches the management plane | Trade-off |
|---|---|---|
| Full | Request and response bodies, plus metadata | Most detail for review, but prompts leave your network |
| Metadata only | Identifiers, counters, cost, timing, and three routing headers | Prompt and response bodies stay in your cluster |
| Off | Nothing | The management plane keeps no request record. New usage and spend go uncounted, so a budget can let requests through past its limit |
The data plane still sends the full record to any logging destination you run yourself, whatever the mode. A team can keep prompts out of the management plane and still send full records to its own SIEM. Each mode change writes an audit log entry, so the auditor can see which mode applied during the review period.
Audit evidence starts at the gateway
An AI gateway turns guardrails into records an auditor can sample. The audit log shows who changed each control. Policy decision records show that the control ran on each request. Mapped to your SOC 2 or ISO 27001 controls and exported to storage you protect, these records answer the auditor’s question with evidence from the review period. Application controls still need their own evidence, so plan for both layers.
Tetrate Agent Router Enterprise is built on the Agent Router project (formerly Envoy AI Gateway), an open source data plane at the Agentic AI Foundation (AAIF) that runs on the Envoy proxy. Agent Router Enterprise runs gateways in any cloud, region, on-premises site, or edge location, and manages all of the gateways from one management plane.
Start a fast-track evaluation: first routed request in about 15 minutes after onboarding, nothing to install. Request a demo to see the audit log and policy decision records in your environment, or review pricing options for enterprise deployment.
Agent Router Enterprise
FAQs
Which AI guardrail evidence comes from the gateway, and which from the application?
The gateway produces policy decision records for every request it handles and an audit log of administrative changes. The application produces evidence for business-logic checks, such as approval limits, that the gateway can’t see. Egress rules cover traffic that tries to bypass both.
Are AI gateway logs immutable?
That depends on the stream. In Agent Router Enterprise, no user can edit or delete an audit log entry, and signed Merkle checkpoints make tampering with sealed windows of the log detectable. Policy decision records get their protection from the place you export them to, so set write-once storage there.
How long should AI guardrail records be kept?
Keep them for at least the full review period, plus any longer period your own policy or a regulation requires. Exported copies follow the destination’s retention policy, so set retention in your SIEM or storage service with your compliance team.
Does Tetrate’s SOC 2 report cover my AI controls?
Tetrate’s SOC 2 Type II report and ISO 27001 certification cover the controls Tetrate operates. Review the reports in Tetrate’s Trust Center as part of your vendor review. The controls you configure, such as guardrail rules and retention, need evidence from your own records.
Key terms glossary
AI guardrails: Runtime rules that allow, block, modify, or redact AI requests and responses. For an audit, a guardrail counts only when it leaves a record that shows it ran.
AI gateway: The routing layer every model call and tool call passes through. Because all traffic shares one path, the gateway can write one record format for every request.
Control evidence: A record from the review period that shows a control operated. Auditors sample control evidence to test operating effectiveness.
Audit log: The record of administrative actions in Agent Router Enterprise, such as key and guardrail changes. No user can edit or delete an entry.
Policy decision record: One verdict per request, with the correlation ID, policy version, caller identity, model, and action taken.
Correlation ID: The identifier that ties a policy decision record to one request. An auditor uses it to trace a sampled request through your logs.
Trust Services Criteria: The criteria a SOC 2 report tests controls against, such as CC6.1 for logical access.
Annex A: The list of 93 reference controls in ISO/IEC 27001:2022, such as A.8.15 for logging.
SIEM: A security information and event management system that collects and keeps log data. Both Agent Router Enterprise record streams can export to one.
Learn more about Tetrate Agent Router Enterprise — enterprise AI agent routing with smart routing, policy, cost controls, and audit across every gateway.