SOC 2 Type II for AI agents
2026 SOC 2 questionnaires carry 15–20 agent-specific items: prompt logging retention, model versioning, sub-processor LLM chain, hallucination handling, agent decision lineage. Generic CC7 logging does not satisfy these.
Mapping Ujex to the agent-specific SOC 2 items
| SOC 2 agent item | Ujex source |
|---|---|
| Prompt logging retention ≥ 12 months | Audit payload.action includes prompt-bearing actions; retention default 12mo |
| Model versioning | aat.agent_version on every row |
| Sub-processor LLM chain | meta.model + meta.vendor stamped by Budgets preflight |
| Agent decision lineage | aat.parent_record_id chains approve → act → write |
| Hallucination handling | Audit row on every overridden / corrected output; Approvals decisions linked |
| Tenant isolation | Pods + Security Rules ownership check |
| Tool authorization | Signed tool manifests, server-side policy, server-injected credentials, and Budgets preflight |
Audit evidence
Use auditExportAATBundle over the auditor's window. The bundle includes the chain head hash so the auditor can verify completeness independently using ujex-audit-chain.
What's NOT covered
- Annual penetration test
- Vendor risk management on your own LLM provider (OpenAI / Anthropic / GCP)
- Background checks / personnel security
- Physical security of your on-prem deployment
Ujex provides the technical artefacts the agent-specific SOC 2 items demand; the rest of the SOC 2 program is the operator's.