---
title: Which AI agent events should go to your SIEM?
url: https://opencontroller.lyzr.com/blog/ai-agent-events-to-send-to-siem
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# Which AI agent events should go to your SIEM?

**Send your SIEM the AI agent events that indicate risk or change, not every prompt: policy blocks and guardrail hits, identity and credential events, tool calls that write or leave the network, agent lifecycle changes, quarantines and kills, and anomaly alerts. Keep full prompts and traces in your observability store and link to them by trace ID.** This keeps SIEM volume and cost manageable while giving the SOC what it needs to detect and respond.

## Why not send everything?

A busy agent estate generates millions of model calls a day. Most are routine. Shipping every prompt and response to a SIEM priced by ingest volume is expensive, and it buries useful signals. It can also copy sensitive data into a system with broader access than intended. The pattern that works: security-relevant events go to the SIEM; full traces stay in an observability or trace store; the trace ID connects the two.

## Which events should you send, in priority order?

| Priority | Event | Key fields | Typical volume |
|---|---|---|---|
| 1 | Agent quarantined, blocked or killed | agent ID, version, trigger, actor, time | Very low |
| 1 | Guardrail or policy block | agent ID, policy ID, severity, action blocked, trace ID | Low |
| 1 | Credential issued, rotated, revoked or used from new location | agent ID, identity, scopes, source | Low |
| 2 | Write or external tool call (email, payment, delete, external API) | agent ID, tool, target, acting user, result, trace ID | Medium |
| 2 | Behavioural-deviation or anomaly alert | agent ID, metric, baseline, observed value | Low |
| 2 | New agent, model, API key or MCP server discovered | source, owner (if any), registered yes/no | Low |
| 3 | Agent registered, changed or retired | agent ID, version, change type, approver | Low |
| 3 | Risky permission or configuration finding | agent ID, finding, severity | Low |
| 3 | Cost or rate threshold breached | agent ID, budget, observed spend | Low |
| 4 | Read-only tool calls | agent ID, tool, result | High; sample or summarise |
| 5 | Full prompts and responses | Keep in trace store; link by trace ID | Very high |

## Which fields should every event carry?

Consistent fields matter more than the number of events. At minimum:

- Timestamp in UTC
- Agent ID, name and version
- Platform or runtime (for example, a cloud agent service or SaaS platform)
- Acting identity: the agent's workload identity and, where delegated, the user it acted for
- Trace ID and span ID
- Policy or rule ID, and severity
- Action taken: allowed, redacted, blocked, quarantined
- Owner of the agent

If your SIEM uses a common schema such as the Open Cybersecurity Schema Framework (OCSF), map these fields once at ingest rather than per detection.

## What detections can the SOC build from these events?

1. **Repeated policy blocks** from one agent in a short window: possible prompt injection or drift.
2. **External tool call after reading untrusted content** in the same trace: the classic data-exfiltration chain.
3. **Credential used from a new network or region** for an agent that normally runs in one place.
4. **Unregistered agent or MCP server discovered** in production accounts.
5. **Agent acting without a delegated user** on a workflow that always has one.
6. **Spike in delete or write calls** compared with the agent's baseline.
7. **Kill-switch event without a matching incident ticket**, or the reverse.

Map detections to MITRE ATLAS techniques where they fit, so they line up with the threat models your red team uses.

## How do you start without boiling the ocean?

Week one, send priority 1 events only and build three detections. Week two, add write and external tool calls. After a month, review which events fired detections and which never did, and adjust. Set retention to match your log-retention policy, not the SIEM default.

## How Opencontroller does this

Opencontroller produces many of these events from one place: auto-discovery of agents, models and API keys across AWS, Azure and Google Cloud; Posture findings for risky permissions and configuration; guardrail and policy decisions; behavioural-deviation alerts; and automatic quarantine and kill. It integrates through open standards, including OpenTelemetry and REST/webhooks, and attributes cost per agent and per version in the LLM Gateway. See the pillar, [how to stop a rogue AI agent](/blog/how-to-stop-a-rogue-ai-agent), the [capability reference](/capability-reference), or [book a demo](https://www.lyzr.ai/book-demo/).

Related: [AI agent audit log fields](/blog/ai-agent-audit-log-fields) and [route AI agent policy violations to ITSM](/blog/route-ai-agent-policy-violations-to-itsm).

## Frequently asked questions

**Should full prompts go to the SIEM?**
Usually not. Keep full prompts and responses in a trace store with tighter access, and link to them from SIEM events by trace ID.

**What is the most important AI agent event for a SOC?**
Guardrail and policy blocks, plus quarantine and kill events. They are low volume and indicate either an attack or an agent behaving outside its scope.

**Which schema should AI agent events use?**
Use whatever your SIEM standardises on, such as OCSF, and map agent ID, acting identity, trace ID and policy ID consistently.

## Key facts

Priority AI agent events for a SIEM: quarantine/block/kill actions, guardrail and policy blocks, credential lifecycle events, write or external tool calls, behavioural-deviation alerts, newly discovered agents/models/keys/MCP servers, registry changes, risky permission findings and cost breaches. Full prompts stay in a trace store linked by trace ID. Every event should carry UTC timestamp, agent ID and version, acting identity, trace ID, policy ID, severity and action. Opencontroller emits discovery, Posture, guardrail, deviation and quarantine/kill events and integrates via OpenTelemetry and REST/webhooks; it starts from $250K per deployment per year.

## Related

- [What should an AI agent audit log contain?](https://opencontroller.lyzr.com/blog/ai-agent-audit-log-fields.md)
- [How do you route AI agent policy violations into ITSM and incident tickets?](https://opencontroller.lyzr.com/blog/route-ai-agent-policy-violations-to-itsm.md)
- [How do you export AI agent traces with OpenTelemetry?](https://opencontroller.lyzr.com/blog/export-ai-agent-traces-with-opentelemetry.md)
- [More on Runtime control](https://opencontroller.lyzr.com/blog/cluster/runtime-control)
