Threat Detections
Threat Detections surfaces suspicious privileged-account behaviour detected in real time by OrbisID's rule-based engine and ML confidence model. Detections are raised when Windows Event Log activity collected by Endpoint Sensors matches a detection rule — for example, a new service being installed, audit policy being modified, or a non-privileged account adding itself to a privileged group.
Requires Enterprise edition — not available on Community.
Threat Detections are produced by Endpoint Sensors, which are optional. Alerts from Microsoft Defender for Identity or CrowdStrike do not appear on this page: they are read under Identity Alerts, shown on the affected account, and counted in its risk score. With neither set up this page is empty, and the threat detection KRIs show Not set up.
Navigating to Threat Detections
Go to Alerts → Threat Detections in the left navigation menu.
Detection Statistics
The four cards at the top of the page give an at-a-glance summary:
| Card | Description |
|---|---|
| Open | Detections that have not yet been actioned |
| Acknowledged | Detections under investigation |
| Confirmed | Detections confirmed as genuine incidents |
| Dismissed | Detections determined to be false positives |
Filtering Detections
Use the Filter by Status dropdown to narrow the list:
- Open (default) — newly raised detections requiring attention
- Acknowledged — detections being investigated
- Confirmed — verified incidents
- Dismissed — closed false positives
Clustered View
Requires OrbisAI with the Threat Detection Clustering feature enabled.
A Clusters | All Alerts toggle appears next to the page title when this feature is on. Clusters group related detections for the same account within a short time window into a single incident card — showing the systems and detection rules involved, the highest confidence score, and a short AI-generated narrative explaining the pattern. Expand a cluster to see and action its individual detections exactly as in the classic list.
Clustering is a deterministic grouping, not a change to detection itself — it never alters which detections are raised or their confidence scores, only how related ones are presented together. If AI is briefly unreachable, clusters still form; only the narrative text is missing. Use the refresh icon to regenerate narratives on demand (distinct from re-running detection). Switching to All Alerts always shows the classic unclustered list below, unaffected by clustering.
Understanding Detection Columns
| Column | Description |
|---|---|
| Account | The account the suspicious activity was observed on |
| System | The target system where the activity occurred |
| Reason | A human-readable explanation of why the detection was raised, including the triggering Event ID |
| Confidence | ML-adjusted confidence score (0–100%). Higher scores indicate stronger evidence |
| Status | Current detection status |
| Detected | When the detection was first raised |
Confidence Score Colouring
| Colour | Range | Interpretation |
|---|---|---|
| Red | 80 – 100% | High confidence — strong anomaly signal |
| Orange | 50 – 79% | Medium confidence — worth investigating |
| Green | 0 – 49% | Low confidence — likely benign or from a known privileged account |
How the Confidence Score Works
Each detection starts with a base confidence determined by the detection rule (e.g., 0.90 for audit policy changes, 0.70 for PowerShell script blocks). The score is then adjusted by:
- Off-hours boost — +0.15 if the event occurred outside 08:00–18:00 on a weekday, or on a weekend
- ML adjustment — a Naive Bayes model trained on your confirmed and dismissed detections adjusts the score based on account name, rule type, PAM risk level, account type (Human/Non-Human), and identity linkage. The model weight scales from 0 → 1 over the first 50 analyst decisions to prevent sparse data from dominating early.
Known privileged service accounts (NON_HUMAN type, HIGH/CRITICAL PAM risk) naturally accumulate false-positive signal, causing the model to lower confidence for expected privileged actions on those accounts over time.
Actioning Detections
For any Open detection, action buttons appear in the Actions column.
Acknowledge (✓)
Mark the detection as under investigation.
- Click the yellow tick icon
- Optionally add investigation notes
- Click Acknowledge
Confirm (shield)
Mark the detection as a genuine incident.
- Click the red shield icon
- Add notes describing the confirmed threat
- Click Confirm
Confirmed detections train the ML model (increasing confidence for similar future events on the same account) and contribute to KRI scoring.
Dismiss (✗)
Mark the detection as a false positive.
- Click the grey X icon
- Optionally explain why it is a false positive
- Click Dismiss
Dismissed detections train the ML model (decreasing confidence for similar future events) and are retained for audit purposes.
Bulk Actions
When multiple detections are selected (using the checkbox column), a bulk action toolbar appears:
- Bulk Acknowledge — acknowledge all selected detections
- Bulk Dismiss — dismiss all selected detections as false positives
- Bulk Confirm — confirm all selected detections as genuine incidents
Bulk confirms and dismisses trigger an immediate ML model retrain.
Viewing Detection Details
Click the eye icon on any row to open the full detection detail dialog. This shows:
- Reason — full description of the triggering event and rule
- Evidence — key–value pairs of supporting data (Event ID, channel, timestamp, process name, object name, logon ID, and raw event data)
- Confidence Score — the current ML-adjusted score
- Endpoint Sensor — which sensor collected the underlying event data
- Notes — any notes added during acknowledgement, dismissal, or confirmation