Processes Defender Scored but Never Alerted On
Medium severity TLP:CLEAR Part 2 of 5 Research by Protokol-7 Published 19 Nov 2026
Hypothesis
An intruder's tooling on a Windows host performed credential access, injection, persistence or defence-evasion behaviour that Microsoft Defender attributed to the process as ATT&CK techniques, but stayed below the threshold for a detection, so nothing reached Protection History or the SIEM.
Brief
Why This Hunt Matters
Microsoft Defender reasons about more than it reports. For each process, its behaviour monitoring keeps a running list of the ATT&CK techniques it has attributed, such as T1003.001 against a renamed procdump or T1547.001 against a rundll32. It raises a detection only when the combined score crosses a threshold. Everything below that line is never shown, in Protection History or anywhere else.
Protokol-7 found that these attributions live in the attack-scoring queues of Defender's encrypted database, mpenginedb.db, keyed by process ID and start time. The open-source osquery-defender-extension decrypts the database and exposes the queues as an osquery table. For the many Windows fleets that run Defender Antivirus without Defender for Endpoint, this is the only behavioural sensor they have, and its sub-threshold judgement was previously thrown away.
How the Hunt Flows
The first query reads the queues on every Windows host. For each process it returns the techniques Defender attributed or that the process inherited from its parent, how many fall in a list of high-signal techniques, and the running process's path and command line where the PID and start time still match. The hunt stacks the image and technique set across the fleet. An updater scored for the same techniques on 3,000 hosts drops away, and a binary in AppData scored for T1003.001 on one host stands out.
Three queries then ask whether Defender said anything about each process. The host's own threat queue shows the behaviour detections the engine named against specific PIDs. The low-fidelity queue shows files sent for a cloud verdict; Protokol-7 found hashcat modules there. Forwarded Defender detections show what reached the SIEM. An agent keeps only the scored processes with no alert of any kind and ranks them.
For those, the hunt rebuilds context from ordinary telemetry: the parent, command line, user and hash of each process, and the files it wrote. A .dmp file written by a process Defender scored for LSASS access settles the question quickly. A second agent gives a verdict per process. Malicious ones go to an approval-gated action that copies the Scans folder and a memory image before isolating the host. Unclear ones go to an analyst.
What the Hunt Cannot See
The technique queues keep entries for about an hour. A live run sees only the last hour; to hunt over days, schedule the extension's defender_mpengine_process_techniques pack query every 30 minutes and run the triage over its results. Without Defender detections in the SIEM, "never alerted" rests on the host's own threat queue alone. Microsoft does not document how techniques are attributed or where the threshold sits, so every row is treated as Defender's opinion, a lead rather than proof. Processes that have exited no longer join to the live process table and need process history from other telemetry.
In this series
Steps
-
Query, baseline
Processes Defender attributed techniques to
Read the attack-scoring queues: for each process, the ATT&CK techniques Defender attributed to it or that it inherited from its parent, keyed by PID and start time, joined to the running process where the PID and start time still match. Defender keeps these entries for about an hour, so this is a view of the last hour unless the pack query has been collecting them.
sql, reads defender_mpengine_rolling_queuesSELECT (SELECT hostname FROM system_info) AS device_hostname, r.pid, r.process_start, CASE WHEN json_valid(r.entry_value) THEN json_extract(r.entry_value, '$.Process_Img') END AS scored_image, COUNT(DISTINCT r.technique_id) AS technique_count, GROUP_CONCAT(DISTINCT r.technique_id) AS techniques, SUM(CASE WHEN instr(',' || LOWER('{{high_signal_techniques}}') || ',', ',' || LOWER(r.technique_id) || ',') > 0 THEN 1 ELSE 0 END) AS high_signal_hits, MAX(CASE WHEN r.queue_type = 'inherited_techniques' THEN 1 ELSE 0 END) AS inherited, MIN(r.insert_time) AS first_scored, MAX(r.insert_time) AS last_scored, p.path AS live_path, p.cmdline AS live_cmdline, p.parent AS live_parent_pidFROM defender_mpengine_rolling_queues rLEFT JOIN processes p ON p.pid = r.pid AND ABS(p.start_time - r.process_start) <= 2WHERE r.queue_type IN ('techniques', 'inherited_techniques') AND r.technique_id != ''GROUP BY r.pid, r.process_startORDER BY high_signal_hits DESC, technique_count DESCWhat a hit looks like. Installers, updaters and management agents are scored every day and appear on many hosts with the same techniques. A user-space image, or a LOLBin with credential-access, injection or persistence techniques on one or two hosts, is the lead. live_path is empty once the process has exited.
-
Query, enrichment
Threats Defender recorded on the host
Read the RecentThreatsOnMachine queue: the behaviour detections the engine named, with the PID and image they were raised against. A scored process with a threat record was noticed; one without was not.
sql, reads defender_mpengine_rolling_queuesSELECT (SELECT hostname FROM system_info) AS device_hostname, entry_key AS threat_name, pid, process_start, CASE WHEN json_valid(entry_value) THEN json_extract(entry_value, '$.Process_Img') END AS image, insert_timeFROM defender_mpengine_rolling_queuesWHERE queue_type = 'recent_threats' AND insert_time >= (CAST(strftime('%s', 'now') AS INTEGER) - {{lookback_days}} * 86400)What a hit looks like. Behavior:Win32/... names against specific PIDs. Matching pid and process_start to scored-processes shows which scored processes Defender did act on.
-
Query, enrichment
Low-fidelity signature hits sent to the cloud
List files that matched a low-fidelity signature and were sent for an asynchronous cloud verdict. These are Defender's maybes; Protokol-7 found hashcat modules here.
sql, reads defender_mpengine_lowfi_asyncSELECT (SELECT hostname FROM system_info) AS device_hostname, file_name, sig_seq, sig_sha, sig_is_sync, instance_timeFROM defender_mpengine_lowfi_asyncWHERE instance_time >= (CAST(strftime('%s', 'now') AS INTEGER) - {{lookback_days}} * 86400)What a hit looks like. A LowFi hit on the scored image, or on a file beside it, strengthens the lead. Signature sequence numbers are not documented.
-
Query, enrichment
Defender detections that reached the SIEM
List the Defender detections forwarded from these hosts in the window, to tell scored processes that alerted from those that did not.
sql, reads hb_detection_findingSELECT device_hostname, time, finding_title, malware_name, file_path, process_path, disposition, severity, metadata_productFROM hb_detection_findingWHERE time >= datetime('now', '-{{lookback_days}} days') AND (LOWER(COALESCE(metadata_product, '')) LIKE '%defender%' OR LOWER(COALESCE(provider, '')) LIKE '%defender%')What a hit looks like. Detections with a process_path or file_path matching a scored image mean the process alerted. No rows from a host with scored processes is either a real gap or detections not being forwarded.
-
Agent triage
Scored but silent
Keep the scored processes nothing alerted on, ranked.
-
Query, enrichment
Process lineage on scored hosts
Recover parent, command line, user and hash for user-space and LOLBin processes on the scored hosts, including ones that have exited, to put each scored PID in its process tree.
sql, reads hb_process_activitySELECT device_hostname, time, pid, parent_pid, parent_process_name, parent_process_cmd_line, process_name, process_cmd_line, user_name, integrity_level, process_hash_sha256FROM hb_process_activityWHERE time >= datetime('now', '-{{lookback_days}} days') AND ('{{scope_hosts}}' = '' OR instr(',' || '{{scope_hosts}}' || ',', ',' || device_hostname || ',') > 0) AND ((LOWER(process_name) LIKE '%\users\%' OR LOWER(process_name) LIKE '%\programdata\%' OR LOWER(process_name) LIKE '%\windows\temp\%' OR LOWER(process_name) LIKE '%\perflogs\%' OR LOWER(process_name) LIKE '%\$recycle.bin\%') OR (LOWER(process_name) LIKE '%\powershell.exe' OR LOWER(process_name) LIKE '%\pwsh.exe' OR LOWER(process_name) LIKE '%\cmd.exe' OR LOWER(process_name) LIKE '%\wscript.exe' OR LOWER(process_name) LIKE '%\cscript.exe' OR LOWER(process_name) LIKE '%\mshta.exe' OR LOWER(process_name) LIKE '%\rundll32.exe' OR LOWER(process_name) LIKE '%\regsvr32.exe' OR LOWER(process_name) LIKE '%\msbuild.exe' OR LOWER(process_name) LIKE '%\installutil.exe' OR LOWER(process_name) LIKE '%\certutil.exe' OR LOWER(process_name) LIKE '%\bitsadmin.exe' OR LOWER(process_name) LIKE '%\msiexec.exe' OR LOWER(process_name) LIKE '%\curl.exe'))What a hit looks like. The scored PID with its parent and command line. An Office, browser or archive-tool parent, or an encoded or download command line, supports the lead.
-
Query, enrichment
Files written by user-space and LOLBin processes
Find the dumps, payloads and scripts these processes wrote. A .dmp written by a process scored for T1003.001 is a credential dump.
sql, reads hb_file_activitySELECT device_hostname, time, process_name, process_cmd_line, file_path, file_hash_sha256FROM hb_file_activityWHERE time >= datetime('now', '-{{lookback_days}} days') AND ('{{scope_hosts}}' = '' OR instr(',' || '{{scope_hosts}}' || ',', ',' || device_hostname || ',') > 0) AND ((LOWER(process_name) LIKE '%\users\%' OR LOWER(process_name) LIKE '%\programdata\%' OR LOWER(process_name) LIKE '%\windows\temp\%' OR LOWER(process_name) LIKE '%\perflogs\%' OR LOWER(process_name) LIKE '%\$recycle.bin\%') OR (LOWER(process_name) LIKE '%\powershell.exe' OR LOWER(process_name) LIKE '%\pwsh.exe' OR LOWER(process_name) LIKE '%\cmd.exe' OR LOWER(process_name) LIKE '%\wscript.exe' OR LOWER(process_name) LIKE '%\cscript.exe' OR LOWER(process_name) LIKE '%\mshta.exe' OR LOWER(process_name) LIKE '%\rundll32.exe' OR LOWER(process_name) LIKE '%\regsvr32.exe' OR LOWER(process_name) LIKE '%\msbuild.exe' OR LOWER(process_name) LIKE '%\installutil.exe' OR LOWER(process_name) LIKE '%\certutil.exe' OR LOWER(process_name) LIKE '%\bitsadmin.exe' OR LOWER(process_name) LIKE '%\msiexec.exe' OR LOWER(process_name) LIKE '%\curl.exe')) AND (LOWER(file_path) LIKE '%.dmp' OR LOWER(file_path) LIKE '%.exe' OR LOWER(file_path) LIKE '%.dll' OR LOWER(file_path) LIKE '%.ps1' OR LOWER(file_path) LIKE '%.bat' OR LOWER(file_path) LIKE '%.vbs' OR LOWER(file_path) LIKE '%.js' OR LOWER(file_path) LIKE '%.hta' OR LOWER(file_path) LIKE '%.zip')What a hit looks like. Files whose writer matches a scored image. Silence is common where file creation is not collected.
-
Agent triage
Scored process verdict
Decide what each silent scored process was.
-
Decision
Route on verdict
Act on malicious processes; send unclear ones to a person.
-
Response action
Isolate the host
Stop the process's activity while preserving evidence.
-
Analyst task
Analyst review
Confirm and extend the finding.
-
Analyst task
Close out
Record results and make the next run see more than an hour.
Coverage
4 of 5 stages covered
| Stage and ATT&CK technique | Covered | How, or why not |
|---|---|---|
| Tool or LOLBin executes T1059.001 T1218.011 |
Yes | scored-processes, scored-process-lineage |
| Credential access behaviour T1003.001 |
Yes | scored-processes, scored-process-file-writes |
| Persistence behaviour T1547.001 T1053.005 |
Yes | scored-processes |
| Injection and evasion T1055 |
Yes | scored-processes, lowfi-cloud-lookups |
| Command and control T1071.001 |
Out of scope | Network connections of the scored process are not examined; pivot to network telemetry for confirmed hosts. |
Blind spots
- Needs the defender_mpengine osquery extension loaded by osqueryd (running as SYSTEM) on every Windows host, with Microsoft Defender Antivirus as the active engine and defender_mpengine_status.error empty. Those hosts return no rows, which looks the same as a clean host. An attacker who has removed or stopped Defender also leaves an empty or stale database. It would answer: Did Defender record anything on hosts where the extension is missing, cannot decrypt the database, or Defender is passive behind another antivirus? Remediation: Deploy the extension through extensions.load on all Windows hosts and alert on defender_mpengine_status rows with a non-empty error.
- Needs the pack query defender_mpengine_process_techniques scheduled every 30 minutes, with its results kept. Technique queue entries expire after about an hour. A live run sees only the last hour, so a burst of activity yesterday leaves no scoring rows. It would answer: What did Defender score more than an hour ago?
- Needs Defender detections (Defender events 1116/1117 or MDE alerts) in hb_detection_finding. Without forwarding, 'never alerted' rests on the host's own threat queue only, and a process can be marked silent when it did alert. It would answer: Did a scored process alert somewhere this hunt cannot see?
- Needs Microsoft documentation of how attack scoring attributes techniques and where the alert threshold sits. The attribution is Defender's opinion and is not documented; the hunt treats it as a lead, never as proof. It would answer: Why was this technique attributed, and how close was it to an alert?
Parameters and data
Parameters
| Parameter | Type | Default | What it is |
|---|---|---|---|
high_signal_techniques | list[string] | T1003, T1003.001, T1055, T1055.012, T1562.001, T1547.001, T1053.005, T1543.003, T1047, T1021.002, T1218.011, T1059.001 | ATT&CK techniques that move a scored process up the list when Defender attributes them. |
lookback_days | number | 14 | Days of history to examine. Defender keeps roughly two weeks in mpenginedb.db. |
scope_hosts | list[host] | none | Hosts to focus the follow-up queries on, as osquery reports their hostname; empty runs them across the estate. |
Data sources
| Source | Category | Telemetry |
|---|---|---|
| Endpoint telemetry (hb_ surfaces) | endpoint | endpoint |
Source
An open hunt.md file. It runs anywhere that reads the format.
---
analysis: The attack-scoring rows are by definition below the threshold Defender alerts
on, and most of them are benign software. Making use of them needs a fleet comparison,
a check against what did alert, and the process tree around each PID, none of which
a rule can supply.
blind_spots:
- id: extension-not-answering
owner: Endpoint engineering
question: Did Defender record anything on hosts where the extension is missing,
cannot decrypt the database, or Defender is passive behind another antivirus?
remediation: Deploy the extension through extensions.load on all Windows hosts and
alert on defender_mpengine_status rows with a non-empty error.
requires: the defender_mpengine osquery extension loaded by osqueryd (running as
SYSTEM) on every Windows host, with Microsoft Defender Antivirus as the active
engine and defender_mpengine_status.error empty
risk: Those hosts return no rows, which looks the same as a clean host. An attacker
who has removed or stopped Defender also leaves an empty or stale database.
- id: scoring-queue-ttl
question: What did Defender score more than an hour ago?
requires: the pack query defender_mpengine_process_techniques scheduled every 30
minutes, with its results kept
risk: Technique queue entries expire after about an hour. A live run sees only the
last hour, so a burst of activity yesterday leaves no scoring rows.
stage: credential-access-behaviour
- id: alert-forwarding
question: Did a scored process alert somewhere this hunt cannot see?
requires: Defender detections (Defender events 1116/1117 or MDE alerts) in hb_detection_finding
risk: Without forwarding, 'never alerted' rests on the host's own threat queue only,
and a process can be marked silent when it did alert.
- id: scoring-semantics
question: Why was this technique attributed, and how close was it to an alert?
requires: Microsoft documentation of how attack scoring attributes techniques and
where the alert threshold sits
risk: The attribution is Defender's opinion and is not documented; the hunt treats
it as a lead, never as proof.
coverage:
- stage: suspicious-tool-execution
status: covered
steps:
- scored-processes
- scored-process-lineage
- stage: credential-access-behaviour
status: covered
steps:
- scored-processes
- scored-process-file-writes
- stage: persistence-behaviour
status: covered
steps:
- scored-processes
- stage: injection-and-evasion
status: covered
steps:
- scored-processes
- lowfi-cloud-lookups
- reason: Network connections of the scored process are not examined; pivot to network
telemetry for confirmed hosts.
stage: command-and-control
status: out_of_scope
guardrails:
claims: no_unsupported
evidence: citation_required
missing_data: not_benign
telemetry: untrusted
hunt:
applicability: environment-specific
handoff: keep-as-periodic-hunt
justification: Defender is the only behavioural sensor on many Windows fleets that
do not license Defender for Endpoint, and its sub-threshold reasoning is otherwise
thrown away.
methodology: structured-hypothesis-driven
trigger: detection-gap
hypothesis: An intruder's tooling on a Windows host performed credential access, injection,
persistence or defence-evasion behaviour that Microsoft Defender attributed to the
process as ATT&CK techniques, but stayed below the threshold for a detection, so
nothing reached Protection History or the SIEM.
labels:
- hunt
- attack.t1003.001
- attack.t1055
- attack.t1059.001
- attack.t1218.011
- attack.t1547.001
- attack.t1053.005
- command and control
- credential access
- defense evasion
- execution
- persistence
- hb-endpoint-control
name: Processes Defender Scored but Never Alerted On
parameters:
high_signal_techniques:
default:
- T1003
- T1003.001
- T1055
- T1055.012
- T1562.001
- T1547.001
- T1053.005
- T1543.003
- T1047
- T1021.002
- T1218.011
- T1059.001
description: ATT&CK techniques that move a scored process up the list when Defender
attributes them.
type: list[string]
lookback_days:
default: '14'
description: Days of history to examine. Defender keeps roughly two weeks in mpenginedb.db.
type: number
scope_hosts:
default: []
description: Hosts to focus the follow-up queries on, as osquery reports their
hostname; empty runs them across the estate.
type: list[host]
provenance:
authors:
- name: Huntbase
org: huntbase.io
generated:
by: hand-authored-with-claude
from: https://protokol-7.github.io/dfir/2026/11/10/defenders-hidden-flight-recorder-mpenginedb.html
gates:
- dry-run
- lint
model: claude-opus-5-5
rationale: 'Run across all Windows hosts, as often as practical: the technique queues
keep about an hour. Where the pack query defender_mpengine_process_techniques has
been scheduled, run the triage over its stored results for the full lookback. Paste
the hosts alert-gap-triage names into scope_hosts before the lineage steps.'
references:
- name: "Protokol-7 \u2014 Defender's hidden flight recorder: mpenginedb.db"
url: https://protokol-7.github.io/dfir/2026/11/10/defenders-hidden-flight-recorder-mpenginedb.html
- name: osquery-defender-extension (defender_mpengine_* tables)
url: https://github.com/rakeshrm/osquery-defender-extension
- name: "Protokol-7 \u2014 mpenginedb_parser (deleted-row recovery, full timeline)"
url: https://github.com/Protokol-7/mpenginedb_parser
related:
- hunt: defender-unsigned-payloads-sideloading
reason: Follows Defender's path tags for the unsigned payloads that scored processes
often start from.
relation: sibling
- hunt: defender-weakening-tamper-asr-exclusions
reason: Checks the engine is running and not weakened, which this hunt depends on.
relation: sibling
scenario:
stages:
- name: Tool or LOLBin executes
observables:
- Tech-pid:<pid>,ProcessStart:<filetime> queues
- Process_Img in a user-writable path
slug: suspicious-tool-execution
tactic: execution
techniques:
- T1059.001
- T1218.011
- name: Credential access behaviour
observables:
- T1003.001 attributed to a non-security process
- .dmp file written
slug: credential-access-behaviour
tactic: credential-access
techniques:
- T1003.001
- name: Persistence behaviour
observables:
- T1547.001 or T1053.005 attributed
slug: persistence-behaviour
tactic: persistence
techniques:
- T1547.001
- T1053.005
- name: Injection and evasion
observables:
- T1055 attributed
- LowFi signature hits
slug: injection-and-evasion
tactic: defense-evasion
techniques:
- T1055
- name: Command and control
observables:
- Outbound connections from the scored process
slug: command-and-control
tactic: command-and-control
techniques:
- T1071.001
summary: Tooling runs on a Windows host and performs credential access, injection,
persistence or defence evasion. Defender attributes ATT&CK techniques to the process
but stays below its alert threshold, so the activity is recorded only in its attack-scoring
queues.
series:
index: 2
slug: defender-mpenginedb-flight-recorder
title: Hunting with Defender's Flight Recorder (mpenginedb.db)
total: 5
severity: medium
targets:
analyst:
name: Tier-2 analyst
role: analyst
endpoint:
category: endpoint
name: Endpoint telemetry (hb_ surfaces)
telemetry:
- endpoint
hb-endpoint-control:
category: endpoint
huntbase:
product: hb-endpoint-control
name: osquery with the Defender MpEngineDb extension (Huntbase Endpoint Control)
hunter:
agent: true
name: Hunt agent
tlp: clear
type: investigation
---
# Processes Defender Scored but Never Alerted On
Defender does more reasoning than it reports. For each process it keeps a running list of the ATT&CK techniques its behaviour monitoring attributed, for example T1003.001 against a renamed procdump or T1547.001 against a rundll32, and it raises a detection only when the score crosses a threshold. Everything below the threshold stays in the attack-scoring queues of mpenginedb.db and is never shown.
This hunt reads those queues on every Windows host through the defender_mpengine osquery extension, joins each PID to the running process, and stacks the image and technique sets across the fleet so that software scored everywhere drops away. It then removes processes Defender did act on, using the host's own threat records, low-fidelity cloud lookups and the detections forwarded to the SIEM. The rest is a list of processes Defender suspected and never mentioned.
Each of those is put back in its process tree and checked for the files it wrote before an agent and an analyst decide what it was. The queues keep about an hour, so the hunt is most useful run often, or over results the pack query has been collecting.
## scored-processes
<!-- Processes Defender attributed techniques to -->
Read the attack-scoring queues: for each process, the ATT&CK techniques Defender attributed to it or that it inherited from its parent, keyed by PID and start time, joined to the running process where the PID and start time still match. Defender keeps these entries for about an hour, so this is a view of the last hour unless the pack query has been collecting them.
```osquery target=hb-endpoint-control role=baseline params=(high_signal_techniques=high_signal_techniques)
~~~yaml
expected: Installers, updaters and management agents are scored every day and appear
on many hosts with the same techniques. A user-space image, or a LOLBin with credential-access,
injection or persistence techniques on one or two hosts, is the lead. live_path
is empty once the process has exited.
prevalence:
by: device_hostname
key:
- scored_image
- techniques
rare_below: 3
reads:
- queue_type
- entry_value
- technique_id
- pid
- process_start
- insert_time
- path
silence: not_evidence_of_absence
source: defender_mpengine_rolling_queues
verified: dry-run
verified_at: '2026-11-19'
~~~
SELECT (SELECT hostname FROM system_info) AS device_hostname, r.pid, r.process_start, CASE WHEN json_valid(r.entry_value) THEN json_extract(r.entry_value, '$.Process_Img') END AS scored_image, COUNT(DISTINCT r.technique_id) AS technique_count, GROUP_CONCAT(DISTINCT r.technique_id) AS techniques, SUM(CASE WHEN instr(',' || LOWER('{{high_signal_techniques}}') || ',', ',' || LOWER(r.technique_id) || ',') > 0 THEN 1 ELSE 0 END) AS high_signal_hits, MAX(CASE WHEN r.queue_type = 'inherited_techniques' THEN 1 ELSE 0 END) AS inherited, MIN(r.insert_time) AS first_scored, MAX(r.insert_time) AS last_scored, p.path AS live_path, p.cmdline AS live_cmdline, p.parent AS live_parent_pid FROM defender_mpengine_rolling_queues r LEFT JOIN processes p ON p.pid = r.pid AND ABS(p.start_time - r.process_start) <= 2 WHERE r.queue_type IN ('techniques', 'inherited_techniques') AND r.technique_id != '' GROUP BY r.pid, r.process_start ORDER BY high_signal_hits DESC, technique_count DESC
```
## parallel-alert-check
<!-- Did Defender say anything -->
parallel:
- → defender-threat-records
- → lowfi-cloud-lookups
- → defender-detections-reported
join: → alert-gap-triage
## defender-threat-records
<!-- Threats Defender recorded on the host -->
Read the RecentThreatsOnMachine queue: the behaviour detections the engine named, with the PID and image they were raised against. A scored process with a threat record was noticed; one without was not.
```osquery target=hb-endpoint-control role=enrichment params=(lookback_days=lookback_days)
~~~yaml
expected: Behavior:Win32/... names against specific PIDs. Matching pid and process_start
to scored-processes shows which scored processes Defender did act on.
reads:
- queue_type
- entry_key
- entry_value
- pid
- process_start
- insert_time
silence: not_evidence_of_absence
source: defender_mpengine_rolling_queues
verified: dry-run
verified_at: '2026-11-19'
~~~
SELECT (SELECT hostname FROM system_info) AS device_hostname, entry_key AS threat_name, pid, process_start, CASE WHEN json_valid(entry_value) THEN json_extract(entry_value, '$.Process_Img') END AS image, insert_time FROM defender_mpengine_rolling_queues WHERE queue_type = 'recent_threats' AND insert_time >= (CAST(strftime('%s', 'now') AS INTEGER) - {{lookback_days}} * 86400)
```
## lowfi-cloud-lookups
<!-- Low-fidelity signature hits sent to the cloud -->
List files that matched a low-fidelity signature and were sent for an asynchronous cloud verdict. These are Defender's maybes; Protokol-7 found hashcat modules here.
```osquery target=hb-endpoint-control role=enrichment params=(lookback_days=lookback_days)
~~~yaml
expected: A LowFi hit on the scored image, or on a file beside it, strengthens the
lead. Signature sequence numbers are not documented.
reads:
- instance_time
- file_name
- sig_seq
- sig_sha
- sig_is_sync
silence: not_evidence_of_absence
source: defender_mpengine_lowfi_async
verified: dry-run
verified_at: '2026-11-19'
~~~
SELECT (SELECT hostname FROM system_info) AS device_hostname, file_name, sig_seq, sig_sha, sig_is_sync, instance_time FROM defender_mpengine_lowfi_async WHERE instance_time >= (CAST(strftime('%s', 'now') AS INTEGER) - {{lookback_days}} * 86400)
```
## defender-detections-reported
<!-- Defender detections that reached the SIEM -->
List the Defender detections forwarded from these hosts in the window, to tell scored processes that alerted from those that did not.
```sqlite target=endpoint role=enrichment params=(lookback_days=lookback_days)
~~~yaml
expected: Detections with a process_path or file_path matching a scored image mean
the process alerted. No rows from a host with scored processes is either a real
gap or detections not being forwarded.
reads:
- time
- provider
- finding_title
- severity
- disposition
- malware_name
- file_path
- process_path
- device_hostname
- metadata_product
silence: not_evidence_of_absence
source: hb_detection_finding
verified: dry-run
verified_at: '2026-11-19'
~~~
SELECT device_hostname, time, finding_title, malware_name, file_path, process_path, disposition, severity, metadata_product FROM hb_detection_finding WHERE time >= datetime('now', '-{{lookback_days}} days') AND (LOWER(COALESCE(metadata_product, '')) LIKE '%defender%' OR LOWER(COALESCE(provider, '')) LIKE '%defender%')
```
## alert-gap-triage
<!-- Scored but silent -->
```agent target=hunter
cite: required
context:
- scored-processes
- defender-threat-records
- lowfi-cloud-lookups
- defender-detections-reported
max_iterations: 4
objective: 'For each scored process, decide whether Defender raised anything about
it: a threat record with the same pid and process_start, a detection on the same
host for the same image near the same time, or a LowFi hit on the image. Keep those
with none. Rank them by high_signal_hits, by rarity of the image and technique set
across the fleet, by whether the image is in a user-writable folder or is a LOLBin,
and by inherited techniques, which mean a scored parent started it.'
success_criteria: A ranked list of host, pid, start time, image and techniques for
scored processes with no alert, citing rows; the hosts to paste into scope_hosts.
tools:
- endpoint
- hb-endpoint-control
```
## parallel-scored-context
<!-- What the scored processes did -->
parallel:
- → scored-process-lineage
- → scored-process-file-writes
join: → scored-process-verdict
## scored-process-lineage
<!-- Process lineage on scored hosts -->
Recover parent, command line, user and hash for user-space and LOLBin processes on the scored hosts, including ones that have exited, to put each scored PID in its process tree.
```sqlite target=endpoint role=enrichment params=(lookback_days=lookback_days, scope_hosts=scope_hosts)
~~~yaml
expected: The scored PID with its parent and command line. An Office, browser or archive-tool
parent, or an encoded or download command line, supports the lead.
reads:
- time
- device_hostname
- pid
- process_name
- process_cmd_line
- process_hash_sha256
- parent_pid
- parent_process_name
- parent_process_cmd_line
- user_name
- integrity_level
silence: not_evidence_of_absence
source: hb_process_activity
verified: dry-run
verified_at: '2026-11-19'
~~~
SELECT device_hostname, time, pid, parent_pid, parent_process_name, parent_process_cmd_line, process_name, process_cmd_line, user_name, integrity_level, process_hash_sha256 FROM hb_process_activity WHERE time >= datetime('now', '-{{lookback_days}} days') AND ('{{scope_hosts}}' = '' OR instr(',' || '{{scope_hosts}}' || ',', ',' || device_hostname || ',') > 0) AND ((LOWER(process_name) LIKE '%\users\%' OR LOWER(process_name) LIKE '%\programdata\%' OR LOWER(process_name) LIKE '%\windows\temp\%' OR LOWER(process_name) LIKE '%\perflogs\%' OR LOWER(process_name) LIKE '%\$recycle.bin\%') OR (LOWER(process_name) LIKE '%\powershell.exe' OR LOWER(process_name) LIKE '%\pwsh.exe' OR LOWER(process_name) LIKE '%\cmd.exe' OR LOWER(process_name) LIKE '%\wscript.exe' OR LOWER(process_name) LIKE '%\cscript.exe' OR LOWER(process_name) LIKE '%\mshta.exe' OR LOWER(process_name) LIKE '%\rundll32.exe' OR LOWER(process_name) LIKE '%\regsvr32.exe' OR LOWER(process_name) LIKE '%\msbuild.exe' OR LOWER(process_name) LIKE '%\installutil.exe' OR LOWER(process_name) LIKE '%\certutil.exe' OR LOWER(process_name) LIKE '%\bitsadmin.exe' OR LOWER(process_name) LIKE '%\msiexec.exe' OR LOWER(process_name) LIKE '%\curl.exe'))
```
## scored-process-file-writes
<!-- Files written by user-space and LOLBin processes -->
Find the dumps, payloads and scripts these processes wrote. A .dmp written by a process scored for T1003.001 is a credential dump.
```sqlite target=endpoint role=enrichment params=(lookback_days=lookback_days, scope_hosts=scope_hosts)
~~~yaml
expected: Files whose writer matches a scored image. Silence is common where file
creation is not collected.
reads:
- time
- device_hostname
- file_path
- file_hash_sha256
- process_name
- process_cmd_line
silence: not_evidence_of_absence
source: hb_file_activity
verified: dry-run
verified_at: '2026-11-19'
~~~
SELECT device_hostname, time, process_name, process_cmd_line, file_path, file_hash_sha256 FROM hb_file_activity WHERE time >= datetime('now', '-{{lookback_days}} days') AND ('{{scope_hosts}}' = '' OR instr(',' || '{{scope_hosts}}' || ',', ',' || device_hostname || ',') > 0) AND ((LOWER(process_name) LIKE '%\users\%' OR LOWER(process_name) LIKE '%\programdata\%' OR LOWER(process_name) LIKE '%\windows\temp\%' OR LOWER(process_name) LIKE '%\perflogs\%' OR LOWER(process_name) LIKE '%\$recycle.bin\%') OR (LOWER(process_name) LIKE '%\powershell.exe' OR LOWER(process_name) LIKE '%\pwsh.exe' OR LOWER(process_name) LIKE '%\cmd.exe' OR LOWER(process_name) LIKE '%\wscript.exe' OR LOWER(process_name) LIKE '%\cscript.exe' OR LOWER(process_name) LIKE '%\mshta.exe' OR LOWER(process_name) LIKE '%\rundll32.exe' OR LOWER(process_name) LIKE '%\regsvr32.exe' OR LOWER(process_name) LIKE '%\msbuild.exe' OR LOWER(process_name) LIKE '%\installutil.exe' OR LOWER(process_name) LIKE '%\certutil.exe' OR LOWER(process_name) LIKE '%\bitsadmin.exe' OR LOWER(process_name) LIKE '%\msiexec.exe' OR LOWER(process_name) LIKE '%\curl.exe')) AND (LOWER(file_path) LIKE '%.dmp' OR LOWER(file_path) LIKE '%.exe' OR LOWER(file_path) LIKE '%.dll' OR LOWER(file_path) LIKE '%.ps1' OR LOWER(file_path) LIKE '%.bat' OR LOWER(file_path) LIKE '%.vbs' OR LOWER(file_path) LIKE '%.js' OR LOWER(file_path) LIKE '%.hta' OR LOWER(file_path) LIKE '%.zip')
```
## scored-process-verdict
<!-- Scored process verdict -->
```agent target=hunter
cite: required
context:
- alert-gap-triage
- scored-process-lineage
- scored-process-file-writes
max_iterations: 4
objective: For each process from alert-gap-triage, decide whether it is malicious,
suspicious or benign using its lineage, command line, files written and the techniques
Defender attributed. A signed vendor binary in Program Files scored for the same
techniques on many hosts is benign; a process that wrote a .dmp after a T1003.001
attribution is malicious.
success_criteria: A verdict per scored process citing rows from each step, and the
image and technique pairs judged benign.
tools:
- endpoint
- hb-endpoint-control
```
## route-on-verdict
<!-- Route on verdict -->
if~: "scored-process-verdict finds at least one malicious scored process" (confidence: high, judge=hunter)
then: → isolate-host
indeterminate: → analyst-review
unavailable: → analyst-review (blind_spot: scoring-queue-ttl)
else: → close-out
## isolate-host
<!-- Isolate the host -->
```action target=endpoint
~~~yaml
approval: required
~~~
Copy C:\ProgramData\Microsoft\Windows Defender\Scans and take a memory image first. Then isolate the host, and terminate the scored process if it is still running.
```
→ analyst-review
## analyst-review
<!-- Analyst review -->
```manual target=analyst
Decrypt the copied Scans folder with the extension's --decrypt mode or Protokol-7's mpenginedb_parser to see the full queue history, including deleted rows. Review the accounts that ran the scored process and their other hosts, and check the scored image's hash across the estate.
```
→ close-out
## close-out
<!-- Close out -->
```manual target=analyst
Record the scored processes and verdicts. Schedule the pack query defender_mpengine_process_techniques every 30 minutes, because the queues keep about an hour, and keep the benign image and technique pairs as a suppression list for the next run.
```
→ end
Run this hunt against your own data.
In Huntbase, Scout runs every step across the tools you already connect, checks back through your history and returns a verdict with the evidence. The hunt.md file is open, so you can also take it anywhere that reads the format.
Drafted by machine (hand-authored-with-claude, claude-opus-5-5), checked by dry-run, lint, then reviewed by a person.