Audit & Compliance AI Agent Governance - Board Oversight Briefing

Executive Summary

Quarterly posture for the Audit & Risk Committee - business context, trends, and required actions.

During Q2 2026 (reporting window ending 08 July 2026), the organization observed 10 distinct AI agents across seven LLM gateway and tracing platforms. Compared with Q1 2026, the estate grew by 2 agents (8 to 10). Five agents are enterprise-governed (built or approved internally); five operate from outside our direct control (3 under partner contracts, 2 with no contract on file).

Third-party exposure increased over the quarter: unregistered 'shadow' agents rose from 1 in Q1 2026 to 2 in Q2 2026 (+1). These agents were discovered in production logs but never passed through procurement or security onboarding. Neither shadow agent has been assigned a security classification yet, meaning we cannot state whether their behavior is acceptable, requires monitoring, or should be blocked.

Visibility into the estate improved materially: monitoring coverage reached 7 of 7 platforms in Q2 2026, up from 5 in Q1 2026. We can now see agent traffic on every major gateway (LiteLLM, Portkey, Cloudflare, OpenTelemetry, Helicone, Langfuse, LangSmith). However, log data is loaded on a batch schedule; as of 08 July 2026, no agent showed activity in the trailing 24 hours, so 'active vs dormant' signals may reflect ingest timing rather than actual business usage.

Governance processes have not kept pace with discovery. Three partner agents remain properly registered, but the two shadow agents lack identity verification, contractual cover, and security review. Remediation requires: (1) completing the security-classification workflow described in Appendix A so each log event receives a clear verdict; (2) issuing a registration policy requiring external agents to be declared before production use; (3) increasing ingest frequency so the board receives timely activity signals.

Agent Estate Dashboard

Current = Q2 2026 snapshot. Prior = Q1 2026 board baseline. Trend shows direction between the two quarters.

* The activity data was loaded in a single quarterly batch that is now ~36 days old. Because the "active in 24 hours" test measures the trailing day, every agent currently reads as inactive. This is a data-collection timing effect, not evidence that AI agents stopped operating.

Governance Spotlight

Every agent in scope falls into one of three governance categories. Enterprise-governed agents are built or explicitly approved by our teams. Partner agents are external but registered. Shadow agents appeared in production logs without registration; they represent the highest oversight gap.

Identity assurance levels (what Status means)

Shadow agents - required remediation steps

• Step 1 - Identify owner: determine which business unit triggered each agent.
• Step 2 - Register or retire: add a partner registration record, or disable if unauthorized.
• Step 3 - Security review: complete the Appendix A workflow (cleared / monitor / block).
• Step 4 - Report back: record the decision so next quarter's pack shows a verdict.

Appendix A - Security classification in Discovery

All 280 observed events are currently unreviewed (no verdict recorded). A security review assigns each a plain-language outcome:

Worked example (plain language)

An external 'research assistant' agent appears in our logs and a safety filter flags one request. Today it is 'pending' (risk unknown) and counts toward the review backlog. After a security analyst reviews it, the outcome is CLEARED (behaviour acceptable - no action) or BLOCKED (agent disabled and removed). Next quarter's pack then shows a clear verdict instead of 'pending'.

Recommended Q3 operationalization roadmap

Glossary: Shadow agent = external AI found in logs without registration. Partner agent = external AI with a declared vendor identity. SOC = Security Operations Center (the team that reviews security alerts). Program approvals and budget decisions are presented in the CISO executive pack.