Snippets

KQL snippet: find risky tool actions from an AI agent

A short query to isolate write-capable tool calls, their execution identity and correlation before disabling an AI agent.

24 Jun 2026 aiagentopskqllogsidentityautomationrunbookmonitoringgovernance

When an AI agent appears to have executed the wrong action, first isolate write-capable tool calls and their execution identity, not only the user prompt.

kusto ai-agent-risky-tool-actions.kql

let Window = 6h;
let AgentName = "ops-assistant";
AppTraces
| where TimeGenerated > ago(Window)
| where tostring(Properties.agent_name) == AgentName
| where tostring(Properties.tool_action) in ("create", "update", "delete", "execute", "approve")
| project TimeGenerated,
        Agent=tostring(Properties.agent_name),
        Tool=tostring(Properties.tool_name),
        Action=tostring(Properties.tool_action),
        Target=tostring(Properties.target_resource),
        Actor=tostring(Properties.execution_identity),
        Approval=tostring(Properties.approval_id),
        Source=tostring(Properties.source_reference),
        OperationId
| order by TimeGenerated desc

Quick read: an action without Approval should be treated as a guardrail gap; an unexpected identity points to execution scope; an empty or unapproved source points to grounding and human validation.