SOC 2 compliance for AI agent deployments

SOC 2 is an audit framework based on the Trust Services Criteria (TSC) from the AICPA. When you deploy AI agents as part of your service, auditors will ask how you control, monitor, and log the agent's actions. This guide maps SOC 2 requirements to AI agent safety controls.

Relevant Trust Services Criteria

Security (CC6): Logical and physical access controls

AI agents need access controls just like human users. Document:

• What tools each agent has access to
• What credentials the agent uses
• How agent permissions are provisioned and revoked
• How you prevent privilege escalation

Policy-based tool authorization provides the access control layer. Each policy documents what tools are allowed and under what conditions. Policies are versioned and changes are tracked.

Availability (A1): System monitoring

Demonstrate that you monitor agent operation and can detect failures:

• Behavioral monitoring with Sentinel tracks action rates, denial rates, and anomalies
• Alerts notify operators of unusual patterns
• Kill switches allow terminating agent sessions

Processing integrity (PI1): Complete and accurate processing

Show that agent actions are executed correctly and completely:

• Policy evaluation ensures only authorized actions execute
• Hash-chained receipts provide a tamper-evident record of all actions
• Content scanning prevents processing of malicious or corrupted input

Confidentiality (C1): Protection of confidential information

Demonstrate controls over sensitive data the agent accesses:

• Policy rules restrict which data the agent can read
• Aegis scans for PII and credential exposure
• Output filtering prevents sensitive data from leaving the system

What auditors will ask

"How do you control what the AI agent can do?" Show them your policy files and explain the policy engine. Demonstrate that unauthorized actions are blocked.

"How do you log agent activity?" Show them the receipt chain. Demonstrate chain verification. Explain retention policies.

"How do you detect if something goes wrong?" Show them Sentinel dashboards. Explain EWMA and CUSUM anomaly detection. Show alert routing.

"How do you manage changes to agent permissions?" Show them policy version history. Explain shadow evaluation for testing changes. Show approval processes for policy updates.

Evidence collection

For SOC 2 audits, you need to provide evidence over the audit period (typically 12 months):

• Policy files with version history and change approvals
• Receipt chain exports showing continuous logging
• Chain verification reports showing integrity
• Sentinel alert history showing monitoring is active
• Incident response records if any events occurred

The control plane API provides exports for all of this data.