KQL Search

Query Details

Agent Content Safety Filter Hits

Query

id: 2b3c4d5e-1111-4aaa-9101-0123456789c1
name: Agent - Content-safety filter hits (hate / sexual / violence / self-harm)
description: |
  Hunts Foundry / Agent Service runs where the Azure AI Content Safety
  filter returned a non-safe result on the prompt or completion. Surfaces
  the four standard harm categories (hate, sexual, violence, self-harm)
  together with the severity reported by the filter, the offending agent,
  the conversation id and the input messages so an analyst can triage the
  prompt that tripped the guardrail.

  Reads the real Foundry telemetry shape: spans land in AppDependencies
  with the property bag in Properties; the content-filter verdict is in
  microsoft.foundry.content_filter.results, and the prompt/response text in
  gen_ai.input.messages / gen_ai.output.messages.
query: |
  let window = 1d;
  AppDependencies
  | where TimeGenerated > ago(window)
  | where isnotempty(Properties["microsoft.foundry.content_filter.results"])
  | extend
      Agent      = tostring(Properties["gen_ai.agent.name"]),
      Model      = tostring(Properties["gen_ai.request.model"]),
      ConvId     = tostring(Properties["gen_ai.conversation.id"]),
      ProjectId  = tostring(Properties["microsoft.foundry.project.id"]),
      Prompt     = tostring(Properties["gen_ai.input.messages"]),
      Response   = tostring(Properties["gen_ai.output.messages"]),
      FilterArr  = todynamic(tostring(Properties["microsoft.foundry.content_filter.results"]))
  | mv-expand Entry = FilterArr
  | extend
      SourceType = tostring(Entry.source_type),
      Blocked    = tobool(Entry.blocked),
      Filter     = todynamic(Entry.content_filter_results)
  | extend
      HateSeverity      = tostring(Filter.hate.severity),
      HateFiltered      = tobool(Filter.hate.filtered),
      SexualSeverity    = tostring(Filter.sexual.severity),
      SexualFiltered    = tobool(Filter.sexual.filtered),
      ViolenceSeverity  = tostring(Filter.violence.severity),
      ViolenceFiltered  = tobool(Filter.violence.filtered),
      SelfHarmSeverity  = tostring(Filter.self_harm.severity),
      SelfHarmFiltered  = tobool(Filter.self_harm.filtered)
  | where HateFiltered or SexualFiltered or ViolenceFiltered or SelfHarmFiltered
       or HateSeverity in ("low","medium","high")
       or SexualSeverity in ("low","medium","high")
       or ViolenceSeverity in ("low","medium","high")
       or SelfHarmSeverity in ("low","medium","high")
  | project
      TimeGenerated, Agent, Model, ProjectId, ConvId,
      HateSeverity, SexualSeverity, ViolenceSeverity, SelfHarmSeverity,
      Prompt, Response
  | order by TimeGenerated desc
tactics:
  - Execution
  - Impact
techniques:
  - T1059
tags:
  - Sentinel-As-Code
  - Custom
  - Foundry
  - AI
  - ContentSafety

Explanation

This query is designed to monitor and identify instances where an Azure AI Content Safety filter has flagged content as potentially harmful in four categories: hate, sexual content, violence, and self-harm. Here's a simplified breakdown of what the query does:

  1. Time Frame: It looks at data from the past day (1d).

  2. Data Source: It examines records from AppDependencies where there is a non-empty result from the content filter.

  3. Data Extraction: For each relevant record, it extracts details such as the agent name, model used, conversation ID, project ID, and the input and output messages.

  4. Content Filter Results: It processes the content filter results to determine if any of the four harm categories were flagged as filtered or have a severity level of low, medium, or high.

  5. Filtering: The query filters out records where any of the categories (hate, sexual, violence, self-harm) were flagged or have a notable severity.

  6. Output: It projects key information like the time the event was generated, agent details, severity of each category, and the prompt and response messages.

  7. Sorting: The results are sorted by the time they were generated, in descending order.

This query helps analysts quickly identify and triage potentially harmful content interactions by providing detailed information about the flagged content and its context.

Details

David Alonso profile picture

David Alonso

Released: June 8, 2026

Tables

AppDependencies

Keywords

AppDependenciesPropertiesAgentModelConvIdProjectIdPromptResponseFilterArrEntrySourceTypeBlockedFilterHateSeverityHateFilteredSexualSeveritySexualFilteredViolenceSeverityViolenceFilteredSelfHarmSeveritySelfHarmFilteredTimeGeneratedExecutionImpactTacticsTechniquesTagsSentinelAsCodeCustomFoundryAIContentSafety

Operators

letagowhereisnotemptyextendtostringtodynamicmv-expandtoboolprojectorder by

Actions