Hunting AI Agent Configuration Drift with Microsoft Sentinel

Four KQL patterns for detecting instruction changes, new MCP servers, ownership changes, and organization-wide sharing

I recently authored and contributed four new Microsoft Sentinel hunting queries for detecting security-relevant configuration drift in AI agents. They have been reviewed, approved, and merged into Microsoft's public Azure-Sentinel repository.

I built the queries around four changes that can materially affect an agent's behavior, access, or exposure: instructions being modified, MCP servers being connected, owners being added, and sharing being expanded to the entire organization. Each modification may be legitimate, but each deserves enough context for a security team to verify that it was expected and authorized.

For a security operations team, the difficult question is often not what does this agent look like now? It is what changed since the last known state?

Microsoft Sentinel's AgentsInfo table provides inventory-style snapshots of AI agents and their associated configuration. That makes it useful for more than posture reporting. By comparing a recent snapshot with an earlier baseline, we can hunt for configuration drift that deserves investigation.

This post walks through four practical hunting scenarios:

1. Instructions changed on a previously published agent
2. A newly observed MCP server on an existing agent
3. An owner added to an MCP-enabled agent
4. Sharing expanded from a restricted scope to the entire organization

The complete hunting queries are available in Microsoft's public https://github.com/Azure/Azure-Sentinel/tree/master/Hunting%20Queries/AI%20Agents. The focus here is the detection design behind them, the KQL patterns they share, and the investigation questions they help answer.

What I contributed

I wrote the four standalone hunting queries discussed in this article and submitted them to Azure/Azure-Sentinel in https://github.com/Azure/Azure-Sentinel/pull/14702:

• AI Agents - Instructions changed on previously published agent
• AI Agents - Newly observed MCP server on existing agent
• AI Agents - Owner added to MCP-enabled agent
• AI Agents - Sharing expanded to organization-wide

The contribution went through several rounds of technical review. Across six commits, I aligned the queries with the unified AgentsInfo schema, added schema-tolerant IdentityInfo enrichment, improved entity mappings, bounded the identity lookback, expanded all owner values, and kept the ATT&CK mappings limited to scenarios where a precise technique could be defended. Repository collaborator v-atulyadav approved the final revision, and the four queries were merged into master on July 20, 2026.

This article explains the detection logic and engineering decisions behind that contribution rather than simply reproducing the final YAML files.

Why current-state queries are not enough

A current-state query can answer questions such as:

• Which agents are published?
• Which agents have MCP servers configured?
• Which agents are shared with the organization?
• Who owns a particular agent?

Those are important posture questions, but they do not tell us whether the state is new. An agent with an MCP server might have been reviewed and approved months ago. The same MCP server appearing for the first time today is a different security signal.

Configuration-drift hunting adds the missing time dimension. Instead of treating a risky-looking property as an event, it compares two states of the same agent and reports only meaningful transitions.

The common detection pattern

I used the same basic time model across all four hunts:

let lookback = 14d;
let recent = 2d;

The latest snapshot observed during the last two days becomes the current state. The latest snapshot from the preceding portion of the 14-day lookback becomes the baseline.

Conceptually, the comparison looks like this:

let CurrentState =
    AgentsInfo
    | where Timestamp > ago(recent)
    | summarize arg_max(Timestamp, *) by AgentId
    | where LifecycleStatus != "Deleted";

let BaselineState =
    AgentsInfo
    | where Timestamp between (ago(lookback) .. ago(recent))
    | where LifecycleStatus != "Deleted"
    | summarize arg_max(Timestamp, *) by AgentId;

CurrentState
| join kind=inner BaselineState on AgentId

Several details matter here:

• arg_max(Timestamp, *) by AgentId selects the latest available state for each agent in the relevant time range.
• The inner join restricts results to agents that exist in both periods. A newly created agent is therefore not automatically treated as configuration drift on an existing agent.
• Deleted lifecycle snapshots are excluded so that a deletion record does not become the effective baseline or current configuration.
• The two-day current window is operationally significant. To retain coverage, these hunts should run within two days of a change.

The 14-day and two-day values are practical defaults, not universal constants. Environments with different ingestion cadence or retention requirements can adjust them, but the current and baseline windows must remain non-overlapping.

Scenario 1: Instructions changed on a published agent

An agent's instructions define its default behavior, persona, and operating boundaries. Changing them can be part of normal development, but it can also weaken restrictions, redirect the agent's behavior, or modify how it uses connected capabilities.

The first hunt compares the current and previous instruction values only when the agent was published in both snapshots:

CurrentState
| join kind=inner BaselineState on AgentId
| where CurrentInstructions != PreviousInstructions
| extend PreviousInstructionsHash = hash_sha256(PreviousInstructions),
         CurrentInstructionsHash = hash_sha256(CurrentInstructions),
         InstructionsLengthDelta = strlen(CurrentInstructions) - strlen(PreviousInstructions)

I deliberately chose to expose hashes and a length delta rather than returning both instruction bodies in plaintext. This confirms that a change occurred without unnecessarily spreading potentially sensitive prompts through query results, exports, or screenshots.

Useful investigation questions include:

• Was the change associated with an approved development or release process?
• Did the agent remain published while the instructions changed?
• Were guardrails, declared tools, permissions, or sharing settings modified around the same time?
• Do audit records identify an expected actor and change path?

The query maps well to an integrity-focused investigation. Its MITRE ATT&CK mapping is T1565.001 (Stored Data Manipulation), but the result is still a hunting lead rather than proof of malicious manipulation.

https://github.com/Azure/Azure-Sentinel/blob/master/Hunting%20Queries/AI%20Agents/AgentsInfoInstructionsChangedOnPublishedAgent.yaml

Scenario 2: A newly observed MCP server

Model Context Protocol servers can extend an agent with external tools, data sources, or actions. From a defender's perspective, the important transition is not simply that an MCP server exists. It is that a server name appears in the current configuration but was absent from the baseline.

The query expands the dynamic McpServers array and builds a set of server names for each agent:

let CurrentMcp =
    CurrentRaw
    | mv-expand Mcp = McpServers
    | extend McpName = tostring(Mcp.name)
    | where isnotempty(McpName)
    | summarize CurrentMcpServers = make_set(McpName) by AgentI

It performs the same normalization for the baseline, then calculates the difference:

| extend AddedMcpServers = set_difference(CurrentMcpServers, BaselineMcpServers)
| where array_length(AddedMcpServers) > 0

Using set_difference() avoids raising a result merely because the order of array elements changed. The hunt reports only MCP server names present in the current set and absent from the previous set.

An analyst should validate more than the displayed name:

• Is the MCP integration part of the approved inventory?
• What endpoint, authentication method, and permissions are associated with it?
• Which tools or data can the server expose to the agent?
• Was the integration introduced through an expected deployment path?
• Did ownership, instructions, or sharing change in the same period?

I did not assign an ATT&CK technique to this query. Adding an MCP server does not, by itself, prove command execution, persistence, or a specific attacker behavior. Avoiding an overly broad mapping keeps the signal honest.

https://github.com/Azure/Azure-Sentinel/blob/master/Hunting%20Queries/AI%20Agents/AgentsInfoNewlyObservedMcpServer.yaml

Scenario 3: An owner added to an MCP-enabled agent

Ownership is a control-plane relationship. A newly added owner may be able to modify an agent's configuration, instructions, integrations, or publication state. The risk becomes more interesting when the agent already has MCP servers configured.

The hunt first limits the current state to MCP-enabled agents:

| where array_length(coalesce(McpServers, dynamic([]))) > 0
| project AgentId, Timestamp, Name, Platform, CreatedDateTime,
          CurrentOwners = coalesce(Owners, dynamic([])), McpServers

It then compares the owner arrays as sets:

| extend AddedOwners = set_difference(CurrentOwners, PreviousOwners)
| where array_length(AddedOwners) > 0
| mv-expand AddedOwnerId = AddedOwners to typeof(string)

Expanding AddedOwners produces one row per newly observed owner. This is more useful than returning one opaque dynamic array because every added identity can be enriched, mapped, and investigated independently.

I kept the raw object identifier in the result even when identity enrichment fails:

| extend AddedOwnerUpn = AccountUpn,
         UnresolvedAddedOwnerId = iff(isempty(AccountUpn), AddedOwnerId, "")

That fallback matters. A missing UPN should not hide the underlying ownership change.

Investigation should establish:

• Is the added owner an expected person, service identity, or administrative group?
• Does the identity's role and business function justify control of this agent?
• Was the owner added before other configuration changes?
• Does the identity appear in related sign-in, audit, or privileged-access activity?
• Should ownership be removed while the change is reviewed?

This query maps to T1098 (Account Manipulation) under Persistence and Privilege Escalation. As with the instruction-change hunt, the mapping frames an investigation hypothesis; it does not label every ownership change as malicious.

https://github.com/Azure/Azure-Sentinel/blob/master/Hunting%20Queries/AI%20Agents/AgentsInfoOwnerAddedToMcpAgent.yaml

Scenario 4: Sharing expanded to the entire organization

An agent can move from a limited audience to organization-wide availability without changing its underlying tools or instructions. That transition can materially increase exposure, especially when the agent has MCP integrations or declared tools.

The hunt treats "*" in SharedWith as the organization-wide state. The current snapshot must contain it, while the baseline must not:

// Current state
| where set_has_element(coalesce(SharedWith, dynamic([])), "*")

// Baseline state
| where not(set_has_element(coalesce(SharedWith, dynamic([])), "*"))

The result also counts MCP servers and declared tools:

| extend McpServerCount = array_length(coalesce(McpServers, dynamic([]))),
         DeclaredToolCount = array_length(coalesce(DeclaredTools, dynamic([])))
| extend HasElevatedCapabilities = McpServerCount > 0 or DeclaredToolCount > 0
| sort by HasElevatedCapabilities desc, Timestamp desc

I use HasElevatedCapabilities as a prioritization field, not a verdict. It brings agents with connected capabilities to the top of the result set so analysts can review the potentially larger blast radius first.

Questions for triage include:

• Was organization-wide publication explicitly approved?
• Is the agent intended for every user, or was a group-based scope expected?
• What data sources, tools, and MCP servers can organization-wide users reach through it?
• Do the instructions contain assumptions that were safe only for a restricted audience?
• Were access reviews or user-acceptance tests completed before the expansion?

No ATT&CK mapping is assigned because a broader sharing scope is a security-relevant exposure change, but not a sufficiently precise adversary technique on its own.

https://github.com/Azure/Azure-Sentinel/blob/master/Hunting%20Queries/AI%20Agents/AgentsInfoSharingExpandedToOrgWide.yaml

Resolving owners without making the hunt schema-fragile

The Owners field contains identifiers. Human-readable identity context makes results easier to triage, and entity mappings make those identities more useful in Sentinel investigations.

I built a small, materialized IdentityInfo lookup that is shared by the four hunts:

let IdentityIdtoUPN = materialize(
    IdentityInfo
    | extend ResolvedAccountUpn = tostring(
        column_ifexists("AccountUpn", column_ifexists("AccountUPN", ""))),
        IdentityTimestamp = todatetime(
        column_ifexists("Timestamp", column_ifexists("TimeGenerated", datetime(null))))
    | where IdentityTimestamp >= ago(lookback)
    | where isnotempty(AccountObjectId) and isnotempty(ResolvedAccountUpn)
    | summarize arg_max(IdentityTimestamp, ResolvedAccountUpn) by AccountObjectId
    | project AccountObjectId = tostring(AccountObjectId),
              AccountUpn = ResolvedAccountUpn);

There are three design choices worth noting:

1. column_ifexists() accommodates observed IdentityInfo naming variants without maintaining separate query versions.
2. The lookup is bounded by the same lookback period instead of scanning unbounded identity history.
3. arg_max() keeps the latest usable identity record for each object ID.

After enrichment, the queries map the account using the UPN components and the Entra object ID:

entityMappings:
  - entityType: Account
    fieldMappings:
      - identifier: Name
        columnName: OwnerAccountName
      - identifier: UPNSuffix
        columnName: OwnerAccountUPNSuffix
      - identifier: AadUserId
        columnName: OwnerId

The strong AadUserId identifier remains valuable even when display information changes. Microsoft Sentinel can use mapped entities in bookmarks and investigation experiences, so mapping the changed owner is more than cosmetic enrichment.

Turning a result into an investigation

These queries intentionally stop at the configuration transition. AgentsInfo tells us that two snapshots differ; it does not necessarily tell us who performed the change, through which interface, or whether the action was authorized.

A practical investigation workflow is:

1. Confirm that the two snapshots represent the expected agent and time period.
2. Review the exact changed property and the agent's current capabilities.
3. Identify the owner or newly added owner through IdentityInfo and Entra ID context.
4. Correlate the transition with the relevant audit source for actor attribution.
5. Check for related changes to permissions, tools, data sources, publication state, and sharing.
6. Validate the change against an approved request, release, or ownership process.
7. Restrict, unpublish, or revert the agent if the exposure cannot be justified.

Expected changes can still be useful findings. Repeated legitimate results may reveal that a deployment process lacks a stable change window, that ownership is managed through noisy automation, or that the hunt's timing needs to be aligned with release activity.

Tuning the hunts for your environment

Before operational use, consider the following adjustments:

• Run cadence: Execute within the two-day current window. A daily cadence provides overlap without blending current and baseline periods.
• Lookback: Increase the 14-day lookback only if snapshot history and query cost support it. A longer lookback does not compensate for missing the current window.
• Known change windows: Add watchlists or environment-specific suppression logic for well-controlled automated deployments, while retaining enough context to audit the change.
• Agent scope: Filter by platform, business unit, agent naming convention, or owner if different teams require separate triage queues.
• Risk prioritization: Raise agents with sensitive declared data sources, powerful tools, privileged owners, or broad availability to the top of the result set.
• Audit correlation: Keep attribution logic separate unless the audit source and join keys are stable in your environment. This makes the configuration-drift hunt reusable while allowing each organization to attach its own control-plane evidence.

Test with representative snapshots before treating any hunt as an operational control. In particular, validate array shapes for Owners, McpServers, and SharedWith, confirm the identity fields present in your workspace, and exercise both changed and unchanged states.

Using the queries

The four YAML definitions have been merged into the Hunting Queries/AI Agents folder of Microsoft's Azure-Sentinel repository. Each file contains the complete KQL, description, entity mappings, and ATT&CK mappings where a precise technique applies.

• https://github.com/Azure/Azure-Sentinel/blob/master/Hunting%20Queries/AI%20Agents/AgentsInfoInstructionsChangedOnPublishedAgent.yaml
• https://github.com/Azure/Azure-Sentinel/blob/master/Hunting%20Queries/AI%20Agents/AgentsInfoNewlyObservedMcpServer.yaml
• https://github.com/Azure/Azure-Sentinel/blob/master/Hunting%20Queries/AI%20Agents/AgentsInfoOwnerAddedToMcpAgent.yaml
• https://github.com/Azure/Azure-Sentinel/blob/master/Hunting%20Queries/AI%20Agents/AgentsInfoSharingExpandedToOrgWide.yaml

The broader pattern is reusable beyond these four scenarios: select a stable current snapshot, select a non-overlapping baseline, normalize dynamic properties into comparable sets, calculate the transition, and enrich only after the drift has been identified. That keeps the core detection explainable and gives the analyst the before-and-after context needed for a defensible investigation.

References

• https://learn.microsoft.com/en-us/azure/azure-monitor/reference/tables/agentsinfo
• https://learn.microsoft.com/en-us/azure/azure-monitor/reference/queries/agentsinfo
• https://learn.microsoft.com/en-us/azure/azure-monitor/reference/tables/identityinfo
• https://learn.microsoft.com/en-us/azure/sentinel/entities-reference
• https://github.com/Azure/Azure-Sentinel/pull/14702

I authored the four hunting queries discussed in this article and contributed them to Microsoft's Azure-Sentinel repository as https://github.com/Azure/Azure-Sentinel/pull/14702. The complete implementations and review history are publicly available through the links above.