fixing-google-secops-detections

Tuning and refactoring Google Chronicle Curated Detections to eliminate alert fatigue and fix logic gaps/bugs.

1. O365 Suite Alerting Failure for BRICKSTORM / APT29

Google released a suite of rules to detect bulk email exfiltration from Microsoft 365 Exchange Online by compromised Service Principals (techniques heavily used by APT29/Midnight Blizzard).

The affected curated rules are:

• O365 Mailbox Access by Service Principal with Multiple User Agents
• O365 Multiple Mailboxes Accessed by Service Principal
• O365 Mailbox Access by Service Principal from Multiple ASNs
• O365 Multiple Mailboxes Accessed via Microsoft Graph API

❌ Core Flaws in Google's Logic

This suite of rules contains systemic logic mismatches between the stated objective and the YARA-L implementation, likely due to code reuse across the rule pack. The rules fail to properly distinguish between automated Service Principals and standard human activity, resulting in alerts that trigger on normal employee behavior.

Flaw 1: Grouping Logic Contradiction

Despite the rule description explicitly stating: "Detects a Service Principal with a single O365 session ID...", the YARA-L match section of the "Multiple User Agents" rule groups by $application_id instead of the session ID ($session_id). This completely contradicts the rule's stated objective.

• Impact: Unrelated employee sessions sharing the same enterprise App ID and same shared mail are merged across the entire tenant over a 3-hour window. This triggers false positives whenever any two distinct users authenticate via different user agents (e.g., iPhone vs. Android) under the same application, failing to isolate individual token hijacking.

Flaw 2: Outlook Mobile Public Client Oversight

The rules track anomalous behavioral patterns (changing IPs, ASNs, or User-Agents) tied to a specific ClientAppId. While Google includes an exclusion regex for several native Microsoft applications, they inexplicably omitted the most ubiquitous human mobile client: the Microsoft Outlook Mobile App (27922004-5251-4030-b22d-91ecd9a37ea4).

• Impact: Without filtering out this Public Client, legitimate employees reading shared mailboxes from their smartphones who move between networks (e.g., switching from Wi-Fi to 4G) inherently trigger the Multiple ASNs alerts. Different employees using iOS and Android to check the same departmental mailbox trigger the Multiple User Agents alert.

Flaw 3: Shared Mailbox Collision

To identify backend service accounts, Google's logic relies on this condition: $e.principal.user.userid != $e.target.user.userid

• Impact: This condition simply checks if the actor ID is different from the target mailbox owner. While true for automated service principals, it is also standard behavior for human employees accessing shared or departmental mailboxes (e.g., an operator opening [email protected]). This design flaw generates massive noise on standard operational human traffic.

✅ Tuned YARA-L Solutions for O365 Rules

To fix these design flaws, you have two options depending on your operational needs:

Option 1: Native SIEM Exclusions (Quick Fix) You do not necessarily have to disable the rules or write custom code. You can simply create an Exclusion directly within the Google SecOps SIEM UI. Just add an exclusion targeting the ClientAppId of Outlook Mobile (27922004-5251-4030-b22d-91ecd9a37ea4) and your authorized backup applications (e.g., Keepit). This immediately stops the false positive flood while keeping Google's curated rules active.

Option 2: Deploy Custom Rules (Architectural Fix) If you want to completely fix the underlying grouping logic flaws (such as the $application_id mismatch), you must disable the Curated Detections and deploy Custom Rules. The fixes involve:

1. Explicitly whitelisting Outlook Mobile (27922004-...).
2. Whitelisting known authorized Enterprise Backup Apps (e.g., Keepit).
3. Fixing the match sections to properly aggregate by session.

Tuned Rule 1: Multiple User Agents

Tuned Rule 2: Multiple Mailboxes Accessed

(Same logic, but in the exclusions ensure you whitelist your authorized backup solutions like Keepit (a7cd46df...) alongside Outlook Mobile).

Tuned Rule 3: Multiple ASNs

(Unlike the User Agents rule, Google actually managed to group by $session_id correctly in this one. However, it still lacks the Outlook Mobile exclusion. Follow the same exclusion logic as above, keeping the condition on $source_asn_dc >= 2).

Tuned Rule 4: Multiple Mailboxes Accessed via Microsoft Graph API

(Same logic. Ensure you append authorized backup applications like Keepit (a7cd46df...) to the exclusion regex at the end of the events block).

2. O365 Teams: External Impersonation Attempt

Rule: ttp_o365_teams_external_impersonation_attempt

Target Event: OFFICE_365 (metadata.product_event_type = "TeamsImpersonationDetected")

Microsoft Defender for Office 365 inspects external inbound Microsoft Teams chats and flags potential domain or brand impersonation via the TeamsImpersonationDetected audit operation. When converted into YARA-L detection rules, the native curated implementation suffers from entity misattribution and match-key oversights.

❌ Core Flaws in the Detection Logic

Flaw 1: Adversary vs. Target Inversion

Chronicle's default parser maps the recipient UserId (the internal user who received the message) into principal.user.userid. The external actor attempting the impersonation resides inside the nested Sender object (Sender.UPN), parsed into additional.fields["Sender_UPN"].

• Impact: By assigning $adversary_name = array_distinct($e.principal.user.userid), the rule labels the internal victim as the adversary in downstream SOC alerts and SOAR cases.

Flaw 2: Erroneous Match Grouping Key

The rule enforces match: $principal_userid over 1h, grouping alerts by the recipient employee.

• Impact: If an attacker conducts an external phishing campaign targeting 30 internal employees simultaneously, SecOps generates 30 isolated alerts instead of a single consolidated threat campaign. Conversely, if multiple distinct external actors reach out to the same internal user within 1 hour, their identities are merged into a single alert, obscuring multi-vector attempts.

✅ Tuned YARA-L Solution for Teams Impersonation

Option 1: Alert Profile Adjustment (Quick Fix)

A straightforward operational fix is to disable alerting on the Broad ruleset for this curated detection and keep strictly the Precise channel enabled. This prevents benign cross-tenant or unconfirmed impersonation events from flooding the analyst queue without requiring custom rule deployments.

Option 2: Deploy Custom Rule (Architectural Fix)

This custom rule keeps the original schema intact, modifying only the sender extraction, the match grouping key, and the adversary outcome field:

3. UEBA: Anomalous Auth Attempts Total

Rule: Anomalous Auth Attempts Total by Principal Hostname and Target User ID

(Note: UEBA stands for "User and Entity Behavior Analytics". These rules do not use static signatures, but instead rely on mathematical algorithms to baseline "normal" behavior and alert on statistical deviations).

❌ Hardcoding Flaw & Alert Fatigue

This UEBA rule attempts to detect anomalous authentication spikes by calculating the historical average and standard deviations over 30 days.

• Alert Fatigue: In our production environment, this curated rule generated an incredibly low True Positive rate (~0.08%, with 2 actionable tickets out of 2324 alerts). It is essentially pure noise.
• Code Flaw: Google declares a variable $num_stddevs_away = max(2) at the beginning of the outcome block. However, in the $historical_threshold calculation, the Google developer hardcoded the value 2 instead of using the variable. This programming error prevents analysts from easily overriding the sensitivity via the UI or inherited variables without completely cloning and rewriting the YARA-L logic.

✅ UEBA Tuning Recommendations

Real-world testing shows that simply tweaking statistical thresholds (e.g., raising $num_stddevs_away to 3 or 4, lowering $coefficient_of_variation_threshold from 0.1 to 0.05, or increasing $observation_threshold to 15) is insufficient: it drops counts from 710 to 111 weekly alerts, which remains excessively noisy for an analyst team.

• Recommended Approach: For SecOps enterprise tenants, disable the Broad ruleset alerting for "Failed Authentications by Device" and rely strictly on the Precise alerting channel. This structural mitigation is the only effective way to stop the alert flood.
• Alternative Custom Approach: If you must keep it active, clone the rule into a Custom Rule, fix the hardcoded 2 inside $historical_threshold to match your custom $num_stddevs_away variable, and enforce stricter baseline coefficients.

Disclaimer: These tunings are based on real-world incident response and SIEM engineering experience. Always test YARA-L rules in your specific environment before deploying them to production.