KQL Search

Query Details

Custom Agents - Explicitly recorded guardrail block

Custom Explicit Guardrail Block

Query

let Evidence=materialize(CustomAgentInlineContent(1h));
CustomAgentSpans(1h)
| where Classification !in ('Excluded managed platform', 'Unattributed GenAI') and HasLineage
| where Blocked
| summarize arg_min(EventTime, *) by _ResourceId, OperationId, AgentKey
| extend Signal='Explicit guardrail block recorded; review as a control event'
| extend TimeGenerated=EventTime, ResourceId=_ResourceId, TraceId=OperationId,
    EventCount=tolong(column_ifexists('EventCount', 1)),
    SuccessfulCalls=tolong(column_ifexists('SuccessfulCalls', 0)),
    FailedCalls=tolong(column_ifexists('FailedCalls', 0))
| project TimeGenerated, ResourceId, Agent, AgentId, AgentKey, AppRoleName,
    TraceId, SpanId, Conversation, Model, Tool, TargetHost, ServerDisplay,
    Signal, EventCount, SuccessfulCalls, FailedCalls,
    TriggerSpanId=tostring(column_ifexists('TriggerSpanId', '')),
    TriggerTime=todatetime(column_ifexists('TriggerTime', datetime(null))),
    TriggerTool=tostring(column_ifexists('TriggerTool', '')),
    TriggerToolResult=tostring(column_ifexists('TriggerToolResult', '')),
    TriggerServerDisplay=tostring(column_ifexists('TriggerServerDisplay', '')),
    TriggerTargetHost=tostring(column_ifexists('TriggerTargetHost', '')),
    RiskyOutputKind=tostring(column_ifexists('RiskyOutputKind', '')),
    RelatedToolResults=tolong(column_ifexists('RelatedToolResults', 0)),
    MatchedIntentPrompt=tostring(column_ifexists('IntentPrompt', '')),
    MatchedIntentSpanId=tostring(column_ifexists('IntentSpanId', ''))
| join kind=leftouter (Evidence
    | project ResourceId=_ResourceId, TraceId=OperationId, SpanId, AgentKey, Prompt, Output, ToolArguments, ToolResult, InputMessages, OutputMessages,
    EvidenceSource, PromptEvidenceOrigin, OutputEvidenceOrigin, PromptEvidenceSpanId, OutputEvidenceSpanId,
    InputParseState, OutputParseState, ToolArgumentsState, ToolResultState)
    on ResourceId, TraceId, SpanId, AgentKey
| extend Prompt=coalesce(MatchedIntentPrompt, Prompt),
    PromptEvidenceOrigin=iff(isnotempty(MatchedIntentPrompt), 'Matched intent span: user role', PromptEvidenceOrigin),
    PromptEvidenceSpanId=coalesce(MatchedIntentSpanId, PromptEvidenceSpanId)
| extend PreviewCharacters=toint(min_of(160, max_of(0, (1500 - string_size(tostring(bag_pack(
    'AgentId', AgentId, 'AgentKey', AgentKey, 'TraceId', TraceId, 'SpanId', SpanId,
    'Conversation', Conversation, 'ResourceId', ResourceId, 'Signal', Signal)))) / 12)))
| extend PromptPreview=substring(Prompt, 0, PreviewCharacters),
    OutputPreview=substring(Output, 0, PreviewCharacters)
| extend EvidencePreviewStatus=case(PreviewCharacters == 0, 'Metadata exhausted preview budget; inspect full query results',
    strlen(Prompt) > PreviewCharacters or strlen(Output) > PreviewCharacters, 'Inline evidence; previews shortened; full text in query results',
    isempty(Prompt) and isempty(Output), 'No prompt/output captured or parsed; not a protection verdict',
    'Inline evidence; full recorded text in query results')
| project TimeGenerated, ResourceId, Agent, AgentId, AgentKey, AppRoleName,
    TraceId, SpanId, Conversation, Model, Tool, TargetHost, ServerDisplay,
    Signal, EventCount, SuccessfulCalls, FailedCalls, Prompt, Output, ToolArguments, ToolResult, InputMessages, OutputMessages,
    EvidenceSource, PromptEvidenceOrigin, OutputEvidenceOrigin, PromptEvidenceSpanId, OutputEvidenceSpanId,
    InputParseState, OutputParseState, ToolArgumentsState, ToolResultState,
    PromptPreview, OutputPreview, EvidencePreviewStatus, TriggerSpanId, TriggerTime, TriggerTool, TriggerToolResult,
    TriggerServerDisplay, TriggerTargetHost, RiskyOutputKind, RelatedToolResults
| order by TimeGenerated desc

Explanation

This query is part of a scheduled alert rule in Azure Sentinel, designed to detect and report specific security-related events. Here's a simplified summary:

  1. Purpose: The query identifies and reports instances where a "guardrail block" has been explicitly recorded. This is not an indication of a security compromise but rather a control measure.

  2. Data Sources: It uses data from Application Insights, specifically the "AppRequests" and "AppDependencies" data types.

  3. Frequency and Period: The query runs every hour and looks back over the past 14 days.

  4. Logic:

    • It first collects evidence from a custom agent's inline content over the past hour.
    • It then filters for events where a block has occurred, excluding certain classifications.
    • The query summarizes these events to find the earliest occurrence for each resource, agent, and trace.
    • It extends the data with additional context and prepares a preview of the prompt and output if available.
    • It joins this data with additional evidence to provide a comprehensive view of the event.

  5. Output: The results include detailed information about the event, such as the time it occurred, the resource and agent involved, and any related evidence. It also provides a preview of the prompt and output if applicable.

  6. Alert Configuration:

    • The alert is configured to create an incident for each result.
    • Incidents are grouped based on specific criteria, such as the agent key and resource ID.

  7. Severity and Status: The severity of the alert is set to "Low," and the alert is currently disabled.

  8. Tags and Version: The query is tagged with relevant keywords and is versioned as "1.2.0".

Overall, this query is designed to monitor and report on specific control events related to custom agents, providing insights into potential security controls being triggered.