Agents in the SIEM
Your SOC sees the login. It can't see the agent.
Sending Cortex Agent traces, usage and data access into Microsoft Sentinel.
Kevin Keller ยท kevinkeller.org
Why it matters
Detection and audit already live in your SIEM
Anomaly detection and audit are two of the five controls every agent platform needs. For many Snowflake customers in Europe, the place they live is Microsoft Sentinel.
Agents don't need a new security stack. They need to show up in the one you already run.
The blind spot
The native connector stops at SQL
| Covered | Not covered |
|---|---|
| Logins, queries, grants, roles, users, tables | AI_OBSERVABILITY_EVENTS (agent traces) |
CORTEX_AGENT_USAGE_HISTORY (tokens, credits) | |
ACCESS_HISTORY (objects and columns touched) |
What Snowflake records
Three sources answer the responder's questions
- The trace: every turn as spans, from planning and tool calls to the answer
- The bill: tokens, credits, model and region per request
- The data: which objects and columns each statement read, now with
agents_info
Who asked what, what did the agent do, what did it touch, what did it cost?
The idea
Two paths, one workspace
Keep the native connector. Add a small pipeline for the three AI sources, via the Azure Monitor Logs Ingestion API.
The build
Four steps from blind spot to detection
- Grants:
MONITORon the agent, on a dedicated audit role - Native connector: install the Snowflake solution from Content Hub
- AI pipeline: one Azure provisioning script, one small Python uploader, three custom tables
- KQL: correlate and detect
The privacy decision
Unredacted content is a choice, not a default
- Metadata only: tool names, latency, model, tokens, errors
- With
READ UNREDACTED: prompts, tool inputs and outputs, answers - No role has it by default, not even
PUBLIC - For many European organisations: metadata in the SIEM, full trace kept in Snowflake
Correlation
Three keys tie it together
| Key | Joins |
|---|---|
SESSION_ID | Agent session โ native login and query tables |
QUERY_ID | SQL statement โ objects and columns touched |
REQUEST_ID | Agent request โ tokens and credits |
From an alert on a sensitive table back to the conversation that caused it.
Know the latencies
Not every source is real time
- Agent traces: available quickly
ACCESS_HISTORY: around three hoursCORTEX_AGENT_USAGE_HISTORY: up to a day- Native connector: about two hours
Write detections that tolerate late data.
Before production
What to change first
- Run as the audit role, not
ACCOUNTADMIN, with key-pair or workload identity - Pull incrementally with a high-water mark, on a schedule
- Carry
agents_infoandagent_typeto tell human from agent queries - Cover every agent, and set retention to match the redaction decision
The takeaway
Turn an agent from an invisible user into something your SOC can investigate.
Personal "art of the possible" project, not an official Snowflake or Microsoft solution