Overview of context-aware analytics

Google Security Operations enables you to view telemetry, entity context, relationships, and vulnerabilities as a single detection within your Google Security Operations account. It provides entity contextualization to enable you to understand both the behavioral patterns in telemetry and the context of those impacted entities from those patterns.


  • Surfacing the permissions for an account on which a brute force login is being attempted.
  • Importance of data hosted by an asset that is also the source of outbound network activity.

Customers can use this contextualization for detection filtering, heuristic alert prioritization, triage, and investigation.

Security analysts and detection engineers typically work to craft a detection on a basic pattern of event telemetry (an outbound network connection), creating numerous detections for their analysts to triage. The analysts attempt to stitch together an understanding of what happened to trigger the alert and how significant the threat is.

Context-aware analytics incorporates advanced enrichment capabilities earlier in the detection authoring and execution workflow, enabling you to provide the following additional capabilities:

  • Making relevant context available for heuristic-driven contextual risk scoring of detections at detection execution time rather than at the human triage stage
  • Reducing time spent on triage and manually stitching together information from disparate IT security systems (EDR consoles, firewall/proxy logs, CMDB and IAM context, vulnerability scan results)
  • Enabling analysts and detection engineers to filter out entire clusters of threats that may be expected or represent little/no danger to the enterprise (malware testing in a sandbox environment, vulnerabilities and anomalous activity in a development network with no sensitive data or access, and more)

Writing rules for context-aware analytics

You can use Detection Engine rules to search for entity context data in your Google Security Operations account.

To search for entity context data, complete the following:

  1. Specify a source using either the udm or entity.

    $eventname.[<source>].field1.field2 For an entity context, <source> is 'graph'. For a UDM event, <source> is 'udm'. If omitted, the <source> defaults to udm.

  2. Specify the entity data:

    $e1.graph.entity.hostname = "my-hostname"

    $e1.graph.entity.relations.relationship = "OWNS"

  3. Specify UDM event data. The following statements are equivalent.

    $e1.udm.principal.asset_id = "my_asset_id"

    $e1.principal.asset_id = "my_asset_id"

You can create many of the same types of rules for entity contexts as you would for UDM events, including the following:

  • Multiple event rules

  • Comparing entity contexts to other entity contexts

  • Comparing entity contexts to UDM events

  • Repeated fields in entity contexts

  • Sliding windows

  • Calculating a risk score for detections

Unlike a UDM event, an entity context does not have a specific timestamp. Each entity context record has a time interval, entity.metadata.interval, over which the entity context is valid. This time interval may not be a day boundary and can be any duration.

A UDM event will be correlated with an entity context record only when the timestamp of the UDM event falls within the time interval of the entity context record. If this condition is not met, the UDM and entity are not evaluated for detections. The detection engine implicitly enforces this and you do not need to specify it as a condition in a rule.

  • When comparing UDM events to an entity context with windowing, an entity context represents a constant value over a specified window.
  • If there are adjacent day buckets where the entity context changes its value, Google Security Operations attempts to match on all of the entity context values, and return any and all matches that are found.

Example rules

Searching for entities with administrator context

The following rule searches for entities that are also tied to administrator privileges. It is looking for times when someone with administrator privilege attempted to log in to or log out of the system.

rule LoginLogout {
    ($log_inout.metadata.event_type = "USER_LOGIN" or  $log_inout.metadata.event_type = "USER_LOGOUT")
    $log_inout.principal.user.user_display_name = $user

    $context.graph.entity.user.user_display_name = $user
    $context.graph.entity.resource.attribute.roles.type = "ADMINISTRATOR"

    $user over 2m

    $log_inout and $context

Sliding window example

The following sliding window example is valid.

rule Detection {
    $e1.graph.entity.hostname = $host
    $e2.udm.principal.hostname = $host

    // Using e2 (a UDM event) as a pivot.
    $host over 3h after $e2

    $e1 and $e2

Invalid sliding window example

The following sliding window example is invalid. Entity context cannot be used as a pivot for a sliding window.

rule Detection {
    $e1.graph.entity.hostname = $host
    $e2.udm.principal.hostname = $host

    // Attempting to use $e1 (an entity context) as a pivot. Invalid.
    $host over 3h after $e1

    $e1 and $e2

Login example using outcome section

The following example uses the outcome section to calculate a risk score for the detection.

rule Detection {
    $auth.metadata.event_type = "USER_LOGIN"
    $auth.metadata.vendor_name = "Acme"
    $auth.metadata.product_name = "Acme SSO"
    $auth.target.user.userid = $user
    $auth.target.user.termination_date.seconds > 0

    $auth.metadata.event_timestamp.seconds >

    $context.graph.metadata.vendor_name = "Microsoft"
    $context.graph.metadata.product_name = "Azure Active Directory"
    $context.graph.metadata.entity_type = "USER"
    $context.graph.entity.user.userid = $user
    $context.graph.entity.user.termination_date.seconds > 0

    $user over 15m

    $risk_score = max(
        if ( $auth.metadata.event_type = "USER_LOGIN", 50) +
        if (
            $context.graph.entity.user.title = "Remote" nocase or
            $context.graph.entity.user.title = "Temp" nocase or
            $context.graph.entity.user.title = "Vendor" nocase, 40) +
        if ( $context.graph.entity.user.title = "Legal" nocase, 10)

    $auth and $context

Suspicious process launch example

The following example evaluates UDM event process data against IOC context data stored as an entity context.

rule ProcessLaunch {
    $ioc.graph.metadata.vendor_name = "ACME"
    $ioc.graph.metadata.product_name = "IOCs"
    $ioc.graph.metadata.entity_type = "FILE"
    $ioc.graph.entity.file.sha256 = $hash

    $process.metadata.event_type = "PROCESS_LAUNCH"
    $process.principal.hostname = $hostname
        not $process.target.process.file.sha256 = "" and
        $process.target.process.file.sha256 = $hash

    $hash over 15m

    $ioc and $process

Additional qualifiers for entity context

To create an event variable that uses an entity context, you must provide a <source> after the event name. The <source> must be graph.

The following pattern refers to an entity context:

  • $e.graph.entity.hostname

Note that there are two equivalent methods of referring to a UDM event:

  • $u.udm.principal.asset_id
  • $u.principal.asset_id

You can mix and match all of these qualifiers in rule text. You can use different qualifiers for the same event as well.

Outcome section

Detection engine supports an outcome section that allows you to derive more information from a rule. The logic defined in the outcome section is evaluated against each detection. If a rule generates N detections, each of the N detections may result in a different set of outcomes.

You can find an example rule that uses the outcome section here.

Detailed usage and syntax of an outcome section can be found in this section.

Outcome section and detection deduplication / detection grouping

For rules with a match section, recall that detections are "grouped by" the match variables. This causes detections to be deduplicated, such that one row is returned for each unique set of match variables and time window.

The outcome variables are ignored when doing this deduplication. Thus, if there are two different detections with the same values for the match variables and time window, but with different values for outcome variables, these will be deduplicated and you will see only one detection. This can happen when a detection was created due to late arriving data, for example. Here's an example that illustrates this case.

rule ExampleOutcomeRule {
    $hostname over <some window>
    $risk_score = <some logic here>

This rule results in the following matches:

Detection 1: hostname: test-hostname time window: [t1, t2] risk_score: 10

Detection 2: hostname: test-hostname time window: [t1, t2] risk_score: 73

Because the match variables and the time window are the same for Detection 1 and Detection 2, these are deduplicated and you will see only one detection, even though the outcome variable, risk_score, is different.

What's next

For information about how Google Security Operations ingests contextual data and enriches entities, see How Google Security Operations enriches event and entity data