How AI Agents Expose Gaps in SOC 2 Access and Audit Controls
Summary
Sponsored analysis argues that AI agents can evade the assumptions behind SOC 2 controls, obscuring ownership, authorization, activity, and offboarding. It recommends treating machine accounts as users and aligning agent access with intended tasks.
Key points
- The article argues that SOC 2 criteria can cover AI agents, but do not explicitly require organizations or auditors to treat them as a distinct identity class.
- Agents may be created through OAuth approvals, API keys, or configuration changes without a clear approval or registration record.
- Agents can lack a named owner or use borrowed human credentials, making access reviews unable to reliably identify who or what performed an action.
- Least-privilege permissions show what an agent can access, but not whether its actions match its intended task.
- The article says agent offboarding, MCP server vendor review, and separation of duties between agents can fall outside or weaken existing control evidence.
- It recommends treating machine accounts as users, including agents in system descriptions, and matching their access to their intended purpose.
Article Details
- Topic
- SOC 2 controls and AI agent identity, access governance, and security risks