Incident Investigation

Incident Investigation - Instructions

Purpose

This skill performs comprehensive security investigations on incidents from Microsoft Defender XDR and Microsoft Sentinel. It retrieves incident details, lists alerts, enumerates assets and evidences, and then performs deep investigation on user-selected entities using appropriate tools and specialized skills.

Investigation Flow:

1. Phase 1: Incident Description - Retrieve metadata, alerts, assets, and evidences
2. Phase 1.5: Alert Cohesion & Adjacent-Alert Sweep - Verify the incident's alerts belong together, and sweep for related alerts left outside it
3. Phase 2: Incident Investigation Menu - Ask the user to select the incident assets and entities that should be investigated.
4. Phase 2-A: User Investigation - Follow user-investigation skill workflow
5. Phase 2-B: Device Investigation - Follow computer-investigation skill workflow
6. Phase 2-C: IoC Investigation - Follow ioc-investigation skill workflow for IPs, URLs, Files, Domains, Hashes
7. Phase 3: Post-Investigation Loop - Offer any remaining entities; skip straight to Phase 4 when nothing material remains
8. Phase 4: Final Summary & Verdict Gates - Pass all four verdict gates before writing any classification

📑 TABLE OF CONTENTS

1. Critical Workflow Rules - Start here!
2. Phase 1: Incident Description - Metadata, Alerts, Assets, Evidences
3. Phase 1.5: Alert Cohesion & Adjacent-Alert Sweep - Scope correction before investigating
4. Phase 2: Incident Investigation Menu - Presenting the options
5. Phase 2-A: User Investigation - Using user-investigation skill
6. Phase 2-B: Device Investigation - Using computer-investigation skill
7. Phase 2-C: IoC Investigation - Using ioc-investigation skill (IPs, URLs, Files, Domains, Hashes)
8. Phase 3: Post-Investigation Loop - Offer remaining entities, or skip to Phase 4
9. Phase 4: Final Summary & Verdict Gates - Baseline, alternatives, attribution, negative findings
10. JSON Export Structure - Required fields
11. Error Handling - Troubleshooting guide

⚠️ CRITICAL WORKFLOW RULES - READ FIRST ⚠️

Before starting ANY incident investigation:

1. ALWAYS complete Phase 1 first - Retrieve full incident description before any deep investigation
2. ALWAYS list Sentinel workspaces at the START of Phase 2 - Call list_sentinel_workspaces MCP tool BEFORE presenting the investigation menu
3. ⛔ ALWAYS complete workspace selection BEFORE any investigation - This is a MANDATORY CHECKPOINT:
   • If 1 workspace: auto-select and display to user
   • If multiple workspaces: ASK USER to select and WAIT for response
   • DO NOT proceed to any entity investigation without a workspace selected
4. ALWAYS present extracted entities to user - After workspace selection, ask user which entities to investigate
5. ALWAYS wait for user confirmation - Do not proceed with deep investigation until user selects entities
6. ALWAYS use the correct tools for each entity type:
   • Users → Follow .github/skills/user-investigation/SKILL.md
   • Devices → Follow .github/skills/computer-investigation/SKILL.md
   • IPs/URLs/Files/Domains/Hashes → Follow .github/skills/ioc-investigation/SKILL.md
7. ALWAYS track and report time after each major step
8. ALWAYS filter evidences - Remove internal IPs (RFC1918) and tenant domains from investigation scope. Also remove all public IPs from the devices listed as assets involved in the incident.
   • Validate infrastructure IPs empirically, not by rule. A CIDR range alone is a hypothesis; cross-account prevalence is evidence. Run the IP Prevalence Check before dismissing any IP as shared infrastructure, CGNAT, VPN egress, or a proxy — and state which method you used.
9. ALWAYS defang malicious/suspicious URLs and IPs - NEVER return them as clickable links. Use defang format: hxxps://evil[.]com, 203[.]0[.]113[.]42
10. ⛔ NEVER auto-select a Sentinel workspace when multiple exist - Workspace selection is MANDATORY:
    • ❌ DO NOT select a workspace on behalf of the user when multiple exist
    • ❌ DO NOT switch to another workspace if a query fails
    • ❌ DO NOT proceed with investigation without explicit user selection
    • ✅ If query fails: STOP, report error, ask user to select different workspace
    • ✅ If multiple workspaces: STOP, list all, WAIT for user selection
    • ✅ Only auto-select if exactly ONE workspace exists
11. ALWAYS run the alert cohesion check before Phase 2 - Incidents can fuse unrelated activity. Verify the alerts share entities; if they form disjoint clusters, investigate and verdict each cluster separately. De-correlated ≠ dismissed (see Phase 1.5)
12. ALWAYS establish a baseline before labelling behaviour anomalous - No entity may be described as anomalous, unusual, or compromised without a stated 14–30 day baseline (see Phase 4 Gate 1)
13. ALWAYS pass all four Verdict Gates before writing a classification - Baseline, alternative hypotheses, attribution evidence, negative findings (see Phase 4)

Incident ID Patterns:

⚠️ Sentinel → Defender XDR ID Mapping (Critical):

When an incident is discovered via Sentinel KQL (e.g., SecurityIncident or SecurityAlert tables), its IDs are Sentinel-local and will NOT work with the Triage MCP:

Rule: When querying SecurityIncident for later Triage MCP drill-down, always project ProviderIncidentId alongside IncidentNumber. Use ProviderIncidentId for all GetIncidentById calls.

Date Range Rules:

• Default analysis window: 7 days before current date to current date (Standard)
• Investigation depth options:
  • Comprehensive: 30 days window (for thorough analysis)
  • Standard: 7 days window (default)
  • Quick: 1 day window (for rapid triage)
• Format: ISO 8601 (e.g., 2026-01-17T00:00:00Z to 2026-01-24T00:00:00Z)

IP Prevalence Check

Confirm whether an IP appears across many unrelated accounts in the window. This is the empirical test behind Critical Rule 8.

let TargetIPs = dynamic(["<ip1>","<ip2>"]);
EntraIdSignInEvents
| where Timestamp between (datetime(<start>) .. datetime(<end>))
| where IPAddress in (TargetIPs)
| summarize DistinctAccounts = dcount(AccountUpn), SignIns = count(),
            Days = dcount(bin(Timestamp, 1d)), SampleAccounts = make_set(AccountUpn, 5)
          by IPAddress
| order by DistinctAccounts desc

Run in Advanced Hunting

Feed each result into Gate 1 — an entity whose baseline is mostly singleton IPs behaves differently from one that lives on shared infrastructure.

Report only the counts this query returned. Never supplement, substitute, or "remember" which accounts share an IP — a co-occurrence recalled from cached context is a hypothesis, and writing it into a column labelled as query output converts it into false evidence.

Phase 1: Incident Description

This phase retrieves and presents all incident information. Follow the exact structure below.

1.1 Incident Metadata

Retrieve and list the incident's metadata using GetIncidentById:

1.2 Incident Alerts

🔴 Tool Selection for Alert Retrieval

Use GetIncidentById with includeAlertsData=true to retrieve incident-specific alerts. This returns only alerts correlated to the incident.

⛔ DO NOT use ListAlerts to retrieve alerts for a specific incident. ListAlerts has NO incidentId parameter — it can only filter by createdAfter, createdBefore, severity, status. Calling it returns all tenant alerts (up to page size 10,000), not incident-specific ones. Any unsupported parameter (e.g., incidentId) is silently ignored.

If GetIncidentById(includeAlertsData=true) returns a truncated or excessively large response (e.g., incident has hundreds of correlated alerts from noise sources like Purview IRM or DLP), use RunAdvancedHuntingQuery as the fallback:

// Get alerts linked to the incident's primary user/entity
AlertInfo
| where Timestamp > datetime(<incident_created_minus_7d>)
| join kind=inner (
    AlertEvidence
    | where Timestamp > datetime(<incident_created_minus_7d>)
    | where EntityType == "User"
    | where AccountUpn =~ "<primary_user_upn>" or AccountObjectId == "<user_object_id>"
    | distinct AlertId
) on AlertId
| project Timestamp, AlertId, Title, Severity, Category, AttackTechniques, DetectionSource, ServiceSource
| order by Timestamp asc

This approach bypasses the Triage MCP's alert cap and gives full control over date range and entity filtering.

Alert Fields to Retrieve

For each alert, retrieve:

• Alert name
• Tags
• Severity
• Investigation state
• Status
• Classification (TruePositive / FalsePositive / BenignPositive / unknown)
• Determination (Malware, Phishing, SecurityTesting, NotMalicious, etc.)
• Impacted assets
• Correlation reason
• Detection source
• First activity
• Last activity

🔴 Surface pre-existing determinations. Classification and Determination are often already populated by a prior analyst or by automated investigation. Show them in the alert table and treat them as evidence — an alert already marked notMalicious materially changes the weight of that alert. Do not silently re-litigate a closed determination; either corroborate it or state explicitly why you disagree.

Presentation Rules:

1. Return as a table (exclude the Alert ID column from display — but retain the IDs, Phase 1.5's sweep needs them)
2. Order by last activity date descending
3. Add row numbers starting from 1
4. If more than 30 alerts exist, note this after the table and provide a Defender portal link
5. NEVER calculate and write the total number of alerts
6. Add a Disposition column and fill it in once Phase 2 completes. Every alert must exit the investigation with exactly one of:

No alert may be silently dropped. An alert missing from the final table is an unaccounted-for detection.

1.3 Incident Assets

Retrieve and list ALL assets involved in the incident by type:

Device Assets:

User Assets:

App Assets:

Cloud Resource Assets:

Count assets by type ONLY after retrieving complete lists.

1.4 Incident Evidences

Retrieve evidences classified as malicious or suspicious only:

Processes (Top 10):

• Get ALL malicious/suspicious processes
• Return only the 10 most probable signs of malicious activity (use judgment)

Files (Top 10):

• Get ALL malicious/suspicious files
• Return only the 10 most probable signs of malicious activity (use judgment)

IP Addresses (Top 10, Filtered):

• Get ALL malicious/suspicious IPs
• Filter out RFC1918 internal IPs: 10.x.x.x, 172.16-31.x.x, 192.168.x.x
• Filter out public IPs associated to the devices listed as assets involved in the incident
• Return only the first 10 from filtered list
• DEFANG ALL IPs: When presenting IPs and domains to the user, ALWAYS use defanged format: 203[.]0[.]113[.]42, evil[.]com. NEVER output clickable malicious indicators.

URLs and DNS Domains (Top 10, Filtered):

• Get ALL malicious/suspicious URLs and DNS Domains
• Filter out tenant domain URLs (DNS domains associated with the organization)
• Return only the first 10 from filtered list
• DEFANG ALL URLs AND DNS DOMAINS: When presenting URLs to the user, ALWAYS use defanged format: hxxps://evil[.]com/path, hxxp://malware[.]net. NEVER output clickable malicious URLs.

AD Domains:

• Return ALL malicious/suspicious AD domains (no limit)

For each evidence type: If more than 10 exist, note this after the table and provide Defender portal link.

1.5 Alert Cohesion & Adjacent-Alert Sweep

Incident correlation is fallible in both directions: it can fuse unrelated activity into one incident, and it can leave related alerts outside it. Run both checks before Phase 2 so the investigation scope is correct from the start.

(a) Cohesion Check — do these alerts belong together?

Group the incident's alerts by shared entity (user, device, IP, session/correlation ID) and present the clusters:

• Single cluster (all alerts share ≥1 entity) → proceed normally, one verdict for the incident.
• ≥2 disjoint clusters (zero shared entities) AND >24h gap between them → ⚠️ flag probable spurious correlation. Investigate each cluster independently and issue a separate verdict per cluster in Phase 4.

🔴 De-correlated ≠ dismissed. A cluster found not to relate to the seed alert is never dropped — it either receives its own verdict in this report or is explicitly handed off as a new investigation. Report the correlation defect itself as a finding: it is actionable detection-engineering feedback regardless of whether the underlying activity was malicious.

(b) Adjacent-Alert Sweep — is anything missing?

The incident's alert list is not the full picture of what its entities did. Sweep for alerts that touch an incident entity but were not correlated into the incident:

let IncidentAlertIds = dynamic(["<alertId1>", "<alertId2>"]);   // from GetIncidentById
let IncidentEntities = dynamic(["user1@contoso.com", "203.0.113.42", "HOST-01"]);
let IncidentObjectIds = dynamic(["<userObjectId1>"]);           // more reliable than UPN
let WindowStart = datetime(<incident_created_minus_7d>);
let WindowEnd   = datetime(<incident_last_activity_plus_1d>);
AlertEvidence
| where Timestamp between (WindowStart .. WindowEnd)
| where AccountUpn in~ (IncidentEntities)
    or AccountObjectId in (IncidentObjectIds)
    or RemoteIP in (IncidentEntities)
    or DeviceName in~ (IncidentEntities)
| where AlertId !in (IncidentAlertIds)
| distinct AlertId
| join kind=inner (
    AlertInfo
    | where Timestamp between (WindowStart .. WindowEnd)
) on AlertId
| project Timestamp, Title, Severity, Category, DetectionSource, ServiceSource, AttackTechniques
| order by Timestamp asc

Run in Advanced Hunting

Tool: RunAdvancedHuntingQuery — the default 7d window is well within the 30d AH limit. For windows >30d use mcp_sentinel-data_query_lake and swap Timestamp → TimeGenerated.

⛔ Mandatory — no substitutes. Prior investigations, cached context, or a high-confidence early verdict tell you how to interpret what this sweep returns; none of them tell you what it would have returned. Run it every time and state the row count, including zero.

Report either outcome:

• Alerts found → add them to scope, present in a table, and note that the platform did not correlate them into the incident.
• None found → state it explicitly: "✅ No additional alerts involving incident entities in [window]."
• Query not executed → the result line MUST read "⚠️ Not executed — sweep incomplete". ✅ asserts a zero row count you do not have.

Phase 2: Incident Investigation Menu

⛔ MANDATORY CHECKPOINT: Workspace Selection

This checkpoint MUST be completed before ANY entity investigation can proceed.

Step 2.1: List Sentinel Workspaces

ALWAYS execute this step first, regardless of any other considerations:

list_sentinel_workspaces (MCP tool)

Store the result. This determines the workflow for Step 2.3.

Step 2.2: Present Entity Summary

Show a summary of the incident entities and assets from Phase 1:

• Users (with UPN and display name)
• Devices (with hostname and risk level)
• URLs (defanged)
• IPs (defanged, filtered)
• File hashes
• Domains (defanged)

🔴 Defang every URL, domain, and IP (Critical Rule 9) — hxxps://evil[.]com/path, evil[.]com, 203[.]0[.]113[.]42. Never output clickable malicious indicators.

Step 2.3: Workspace Selection Gate

IF workspace_count == 1:
    - Auto-select the single workspace
    - Display: "Using Sentinel workspace: [NAME] ([ID])"
    - Set SESSION_WORKSPACE_SELECTED = true
    
ELSE IF workspace_count > 1 AND SESSION_WORKSPACE_SELECTED == false:
    - Display all workspaces with Name and ID
    - ASK USER: "Which Sentinel workspace should I run my searches in? Select one or more, or choose 'all'."
    - WAIT for user response
    - Set SESSION_WORKSPACE_SELECTED = true after selection
    
ELSE IF workspace_count > 1 AND SESSION_WORKSPACE_SELECTED == true:
    - Display: "Continuing with previously selected workspace: [NAME] ([ID])"
    - DO NOT ask again

⛔ DO NOT PROCEED PAST THIS POINT WITHOUT A WORKSPACE SELECTED

If SESSION_WORKSPACE_SELECTED == false after Step 2.3, STOP and ask the user to select a workspace.

Step 2.4: Ask User to Select Entities

Ask the user:

"Which assets and entities involved in the incident should be investigated in depth? Please select them by providing their numbers or names, or simply ask to analyze all of them. The more entities you select, the longer the analysis will take."

🔴 DO NOT OFFER OTHER OPTIONS: Only ask the user whether they want to investigate one or more of the incident entities and assets listed above in more depth.

Read the response.

• If they do not want to proceed with the proposed investigations, ask them what they want to do.
• If they want to proceed with one or more of the proposed investigations, continue with Step 2.5.

Step 2.5: Start Investigations

Pre-flight check: Confirm SESSION_WORKSPACE_SELECTED == true before proceeding.

Proceed in accordance with the instructions described below for Phase 2-A, Phase 2-B, and Phase 2-C. When multiple investigation types are selected (users, devices, IoCs) run them in parallel as much as possible.

Evidence Capture Standard (applies to ALL Phase 2 investigations)

Per-event detail is what makes a verdict defensible and a cluster separation provable. For every event cited as evidence, capture:

Why these specifically:

• HTTP method + response size discriminate enumeration from a single UI lookup. A small, single-object read is not reconnaissance no matter which API served it — do not map it to a discovery technique without volume or response-size support.
• Managed/compliant state + risk level are prerequisites for any compromise verdict. An unfamiliar IP on a compliant, managed device with no risk detection is a materially different finding from the same IP on an unmanaged one.
• Correlation / request / session IDs are what make cross-table pivots and the cluster separation in §1.5 provable rather than asserted.
• Success/failure ratio across the whole event set — always record it. High-volume access with zero authentication failures argues strongly against credential guessing or brute force.

🔴 Validate the central artefact at its source. When an alert's premise rests on a specific artefact, query the authoritative table for that artefact directly. Never infer its nature from the alert title.

The platform's own disposition outranks the alert title. An alert named "phish delivered" whose EmailEvents row shows empty ThreatTypes, empty DetectionMethods, and DeliveryAction: Delivered was not a detected phish. Say so, and record the discrepancy itself as a finding.

Phase 2-A: User Investigation

Pre-requisites (MANDATORY)

⛔ VERIFY BEFORE PROCEEDING:

• ✅ SESSION_WORKSPACE_SELECTED == true (workspace explicitly selected by user)
• ✅ SELECTED_WORKSPACE_IDS array is populated with user's selection
• ✅ User has explicitly selected which user(s) to investigate

If any pre-requisite is FALSE: STOP and return to Phase 2.3 Workspace Selection Gate.

User Investigation Workflow

⚡ PARALLEL EXECUTION: When multiple users are selected, execute user investigations in parallel as much as possible.

📦 WORKSPACE CONTEXT: Pass the selected workspace(s) to all child skill invocations:

• Use SELECTED_WORKSPACE_IDS from Phase 2.3 for all Sentinel queries
• If a query fails with table/workspace error: STOP, report error, ask user to select different workspace
• ⛔ DO NOT automatically retry with a different workspace

For EACH user selected by the user:

🔴 REFERENCE THE SKILL FILE: Read and follow the complete workflow defined in:

.github/skills/user-investigation/SKILL.md

Key Steps (summary - see skill file for full details):

1. Get User Object ID from Microsoft Graph
2. Calculate date ranges based on investigation type (Standard/Quick/Comprehensive)
3. Run parallel data collection:
   • Sign-in anomalies (Signinlogs_Anomalies_KQL_CL — note lowercase 'l' in "logs")
   • Sign-in statistics (apps, locations, IPs)
   • Audit log events
   • Office 365 activity
   • Security incidents involving user
   • Identity Protection risk detections
   • MFA and authentication methods
   • Device compliance status
4. IP enrichment for flagged addresses
5. Compile and present findings
6. Generate HTML report (if requested)

DO NOT copy the full workflow here - always read the skill file for the most current instructions.

Phase 2-B: Device Investigation

Device Investigation Workflow

⚡ PARALLEL EXECUTION: When multiple devices are selected, execute device data collection queries in parallel for ALL devices simultaneously. Run Defender alerts, compliance, logged-on users, vulnerabilities, network/process/file events queries concurrently.

For EACH device selected by the user:

🔴 REFERENCE THE SKILL FILE: Read and follow the complete workflow defined in:

.github/skills/computer-investigation/SKILL.md

Key Steps (summary - see skill file for full details):

1. Get Device IDs (Entra Device ID + Defender Device ID)
2. Determine device type (Entra Joined, Hybrid Joined, Entra Registered)
3. Run parallel data collection:
   • Defender alerts for device
   • Device compliance status
   • Logged-on users
   • Software vulnerabilities
   • Network connections
   • Process events
   • File events
   • Automated investigations
4. Compile and present findings

DO NOT copy the full workflow here - always read the skill file for the most current instructions.