AI Audit Trails for SOC 2
Your auditor asks how you track what your AI agents are doing. You need evidence, not a policy document.
ACP writes one audit row for every governed tool call: who made it, which tool, what was decided, and when. This guide maps those rows to SOC 2 trust service criteria and shows how to export them.
The problem
AI agents call tools. They query databases, create records, post messages and read files. Traditional SOC 2 evidence covers human actions such as login logs, access reviews and change management. When an agent acts for a user through a shared API key or service account, the trail stops at the account.
Your auditor sees:
service-account-botqueried the CRM 47,000 times last month- No way to tie those calls to individual users
- No record of which permissions applied
- No record of what the calls contained
What ACP logs
Each governed call produces an audit row. This is an example of the shape. The full field list is in the Logs API reference.
{
"ts": "2026-03-15T14:23:07.142Z",
"sub": "auth0|8f3a2b1c9d4e5f6a",
"userEmail": "alice@example.com",
"tool": "salesforce.query",
"decision": "allow",
"decisionReason": "within policy",
"outcome": "success",
"agentTier": "interactive",
"agentName": "claude-desktop",
"sessionId": "sess_01",
"requestId": "req_01",
"agentChain": [],
"contentScan": {
"piiDetected": false,
"piiTypes": [],
"action": "none"
}
}
The identity fields (sub, userEmail) come from the verified credential. agentName, agentTier and sessionId are reported by the client, so treat them as labels, not proof.
Rows are kept for 365 days on every plan, and without a limit on Enterprise.
If the gateway can’t be reached, interactive sessions proceed and are marked as running ungoverned. Unattended tiers (subagent, background, API) stay blocked. Mention this to your auditor up front.
Mapping to SOC 2 Trust Service Criteria
CC6.1: Logical access security
Requirement: The entity implements logical access security over information assets.
ACP evidence:
- Calls through ACP’s MCP endpoint require a verified JWT from your identity provider (IdP)
- The JWT signature is checked against the IdP’s JWKS endpoint
- Expired, malformed or unsigned tokens are rejected
- Coding-agent hooks authenticate with workspace API keys instead of IdP tokens
- Rejected hook-key attempts are recorded as rate-limited diagnostics (one entry per key prefix and route per minute), not as audit rows
- Rejected MCP tokens return an error to the client and are not written to the audit trail
What to show your auditor: Your IdP configuration, the list of workspace API keys, and sample audit rows for authenticated calls.
CC6.2: User authentication
Requirement: The entity authenticates users before granting access.
ACP evidence:
- ACP works with your existing IdP (Auth0, Okta, Entra ID)
- The user’s identity is carried through the whole tool call
- Multi-factor authentication is enforced at the IdP, not by ACP
- The
subclaim in each row ties back to one person
What to show your auditor: Audit rows with distinct sub values and matching userEmail addresses.
CC6.3: Authorization and access controls
Requirement: The entity authorizes, modifies, and removes access based on job function.
ACP evidence:
- Each tool requires specific scopes
- Scopes are assigned through IdP roles and groups, which you already manage
- A call to a tool the user lacks scopes for is rejected with a JSON-RPC error, code -32004
Insufficient scope. The error data lists therequired,haveandmissingscopes - Each row records the
decisionanddecisionReason
What to show your auditor: Your tool-to-scope mapping, your IdP role assignments, and audit rows where decision is a deny.
CC7.2: Monitoring for anomalies
Requirement: The entity monitors system components for anomalies.
ACP evidence:
- Per-user rate limits
- Content scanning for PII in tool inputs and outputs, recorded in
contentScan - Built-in rules that block SSN and credit card patterns, and SSRF patterns
Private-range SSRF blocking (localhost, 10.x, 172.16-31.x, 192.168.x, 169.254.x, metadata.google.internal) applies to tools the ACP gateway executes. For tool calls governed at the model proxy, only cloud-metadata endpoints are blocked, because localhost is the developer’s own machine. Calls reported by harness hooks are not covered by this rule.
What to show your auditor: Rate limit configuration, content scanning rules, and sample rows where PII was detected.
CC8.1: Change management
Requirement: The entity authorizes, designs, develops, configures, documents, tests, approves, and implements changes.
ACP evidence:
- Scope-to-tool mappings are explicit and reviewable in the dashboard
- Changes to connectors, PII patterns, agents, API keys and scoped tokens are written to the audit log
- The built-in rules can’t be overridden by workspace configuration
What to show your auditor: Your current tool scope configuration, the audit rows for configuration changes, and the built-in rules documentation.
Exporting audit data
Dashboard export
Go to Logs in the dashboard. Filter by date range, user, tool or decision. Export as CSV or NDJSON.
API export
curl -H "Authorization: Bearer gsk_your-api-key" \
"https://api.agenticcontrolplane.com/<workspace>/admin/audit?since=2026-03-01T00:00:00Z&until=2026-03-31T23:59:59Z&limit=1000"
The key needs the admin.audit.read scope. Paging is by time: request the next window with a later since. Details are in the Logs API reference.
There’s no push or webhook delivery to a SIEM today. Poll this endpoint on a schedule, or export CSV from the dashboard, and feed the result into your SIEM’s own ingestion.
Evidence collection checklist
For your SOC 2 audit, collect:
- Identity provider configuration (issuer, audience, JWKS discovery)
- Workspace API keys in use by coding-agent hooks
- Tool scope mappings (which tools require which permissions)
- Sample audit rows: allowed calls, denied calls, PII detection
- Rate limit configuration (per-user limits)
- Content scanning rules (PII patterns)
- Built-in rules documentation (SSN, credit card and SSRF blocking, with the caveats above)
- User access review (which users hold which roles and scopes in your IdP)
- Audit export path (CSV export or Logs API polling into your SIEM)
- Your retention setting and how you handle the ungoverned-call case
How this differs from regular API logging
Regular API logging records that an API was called. An ACP audit row also records:
- Who: the person, taken from the verified credential, not a shared service account
- What was decided: allow, ask or deny, and the reason
- Whether the content was clean: the PII scan result for the call
- The agent chain: if one agent called another, the chain is kept
That is the difference between “someone called the CRM API” and “Alice queried opportunities from Claude Desktop, the call was allowed, and no PII was detected.”
Back to guides · PII detection → · Set up Auth0 →
Running coding agents? ACP records each tool call its hook sees and makes the dangerous shell commands (a force-push, destructive SQL, curl | sh) wait for you. One command, free up to 5 agents:
curl -fsSL https://agenticcontrolplane.com/install.sh | bash
irm https://agenticcontrolplane.com/install.ps1 | iex