Skip to main content
Dithered dot illustration of an antique printer's type case drawer, a grid of wooden compartments with a few holding metal type

What is OCSF (Open Cybersecurity Schema Framework)?

OCSF is an open schema that gives security events from any vendor the same classes, field names, and enum values. How it works, adopters, and trade-offs.

Explainer
Updated 10 min read

Definition

OCSF

Also called Open Cybersecurity Schema Framework, OCSF schema

OCSF (Open Cybersecurity Schema Framework) is an open, vendor-neutral schema that describes security events and findings with the same event classes, field names, and numeric enum values regardless of which product produced them, so one query or detection can read logs from many sources. AWS and Splunk started it, it launched publicly in August 2022, and it has been a Linux Foundation project since November 2024; the current release is 1.9.0.

Current OCSF schema version, released August 3, 2026
1.9.0
Event categories in OCSF 1.9.0, from System Activity to Unmanned Systems
8
Organizations in the August 2022 launch coalition
18
Date OCSF became a Linux Foundation project
Nov 19, 2024

How it works

OCSF describes every event as an instance of exactly one event class, and every event class belongs to one category. The OCSF white paper lists five building blocks: data types and attributes (with arrays), event classes, categories, profiles, and extensions. Objects are the complex data types, such as Process, Device, User, Malware, and File, that event classes are assembled from.

ConstructWhat it isExample in OCSF 1.9.0
CategoryA domain that groups event classesIdentity & Access Management (3)
Event classOne kind of event, with required fieldsAuthentication [3002]
ObjectA reusable group of related attributesuser, src_endpoint, device
ProfileOptional attributes mixed into many classesCloud, Container, Security Control
ExtensionAdditions outside the core schemaWindows, Linux, macOS platform extensions

Each event carries a few required base attributes: class_uid, category_uid, activity_id, severity_id, time, and metadata. The activity_id says what happened inside the class. In the Authentication class, 1 is Logon and 2 is Logoff. The schema then derives type_uid as class_uid * 100 + activity_id, so a logon is 300201. Outcomes are enums too: status_id is 1 for Success and 2 for Failure. Because the numbers are fixed by the schema, a consumer can filter on meaning without knowing which product wrote the event.

OCSF 1.9.0 has eight categories: System Activity, Findings, Identity & Access Management, Network Activity, Discovery, Application Activity, Remediation, and Unmanned Systems. Profiles cut across them. The Security Control profile adds disposition_id, malware, and attacks to activity classes, so an EDR's blocked file open stays a File System Activity event with a "Blocked" disposition instead of needing its own class. Version 1.9.0 also ships an AI Operation profile for model, retrieval, and agent activity. Platform-specific classes live in extensions; Windows Registry Key Activity, for example, is class 201001 in the Windows extension rather than the core schema.

Two base attributes handle what doesn't fit: unmapped holds source fields with no OCSF home, and raw_data can hold the original event. OCSF is agnostic to storage format and ETL, and its definition files are JSON.

Origin, versions, and adopters

OCSF went public at Black Hat USA on August 10, 2022. Splunk's announcement says AWS and Splunk conceived and initiated the project, building on the ICD Schema work done at Symantec, a division of Broadcom. Fifteen more organizations joined the launch, including CrowdStrike, Okta, Palo Alto Networks, and Sumo Logic. On November 19, 2024, OCSF became a Linux Foundation project; the project says its development process didn't change. The schema is licensed under Apache 2.0.

Version 1.7.0 shipped on November 14, 2025, 1.8.0 on March 16, 2026, and 1.9.0 on August 3, 2026.

Adoption shows up in three places:

  • Amazon Security Lake stores data in OCSF. It converts natively supported AWS sources automatically, and custom sources must arrive as OCSF in Apache Parquet. Its version 1 AWS sources are labeled OCSF 1.0.0-rc.2 and its version 2 sources 1.1.0.
  • Splunk ships an ocsf SPL2 command for Edge Processor and Ingest Processor pipelines. It converts _raw to OCSF and prefixes the source type with ocsf: for the OCSF-CIM Add-on and Splunk Enterprise Security.
  • SentinelOne offers Singularity Marketplace connectors that transform third-party data into OCSF for Singularity Data Lake.
  • Cotool's S3 and SQS log ingest offers "Custom format, normalized to OCSF" for sources without a built-in parser, next to its built-in parsers and a Generic JSON option (Cotool docs). The onboarding steps are in the examples below.

OCSF vs ECS, OSSEM, and Splunk CIM

The main difference between these schemas is where normalization happens and who governs the schema. OCSF and ECS rewrite data into a common shape. Splunk CIM and Microsoft's ASIM can leave raw data alone and normalize when you search.

SchemaGoverned byWhen it normalizesBuilt around
OCSFLinux Foundation projectAt ingest or by the producerNumbered event classes and enums
ECSElastic; contributed to OpenTelemetryAt ingest into ElasticsearchField names and Elasticsearch data types
OSSEMOTRF community projectReference for parser authorsData dictionaries, common data model, detection model
Splunk CIMSplunk add-onAt search timeData models of fields and tags
ASIMMicrosoft SentinelQuery-time KQL parsers or ingest-timeNormalized schemas per event type

ECS defines common fields, with Elasticsearch data types, for events stored in Elasticsearch. It is permissive: unmapped fields can be added under custom names. In April 2023 Elastic contributed ECS to OpenTelemetry to merge it with OTel semantic conventions.

OSSEM is a community project by Roberto Rodriguez with three parts: data dictionaries for specific logs, a common data model, and a detection model that maps relationships between events. Microsoft says ASIM aligns with OSSEM.

Splunk CIM is a set of data models applied at search time, a "schema-on-the-fly" that leaves raw data intact. Each data model is a set of fields and tags that define "the least common denominator of a domain of interest." Splunk's OCSF-CIM Add-on supports OCSF-formatted data in Splunk Enterprise Security.

Why it matters for a SOC

A shared schema lets a team write a detection once and run it on every source that maps to it. SentinelOne's OCSF write-up shows one STAR rule on src_endpoint.ip matching SentinelOne, Okta, and Fortinet logs.

Without a shared schema, the same question has a different shape in every tool. A failed Okta sign-in is eventType eq "user.session.start" and outcome.result eq "FAILURE". A failed AWS console sign-in is a CloudTrail ConsoleLogin event with responseElements.ConsoleLogin set to "Failure". In OCSF, both are Authentication events with activity_id = 1 and status_id = 2.

This matters more when an AI agent writes the queries. Against unnormalized logs, an agent has to find each source's field names, sample events to learn what values mean, and guess at equivalents, and any step can fail without an error. With OCSF, the class ID and enums carry the meaning, so the agent can write one filter that applies everywhere. The cost moves to the mapping layer. If a source's mapper never populated status_id, the agent's query silently misses that source. So the agent still needs to know which sources are mapped, and at which OCSF version.

Cotool's federated search shows both halves. Asked which hosts connected to 185.220.101.4 this week, an agent plans one search across four sources and returns three hosts. The firewall and VPC flow logs are stored in Cotool. Splunk, Snowflake, and the CrowdStrike API are queried in place. The stored logs went through Cotool's parsers at ingest. For the other three, the agent translates the question into each system's own fields, which is the per-source work a shared schema removes. That split is a reasonable place to draw the line: normalize the sources queried every day, and query the rest where they are.

The trade-offs: mapping effort and lossy normalization

OCSF moves normalization work from query time to mapping time, and someone still has to build and maintain each mapping.

  • Mapping coverage is per event type, not per vendor. As of November 2026, Splunk Ingest Processor's OCSF conversion supports 12 source types. For OktaIM2:log it converts four event types, and user.session.start isn't one of them. "We support Okta" can mean less than the full feed. Cotool's docs state the same limit for its Generic JSON format: records "are not normalized and no detection coverage is implied."
  • Normalization loses detail. Fields without an OCSF attribute go into unmapped. The schema warns consumers not to rely on that object having a stable structure. raw_data is optional, and the schema says it isn't meant for structured querying. Cotool's log export handles this with two outputs: Normalized, as typed columns in Parquet or JSON Lines, and Raw, the record exactly as the source sent it. Records that fail normalization stay in Cotool, are counted on the export, and are written only to Raw output.
  • Enum choices are judgment calls. severity_id and activity_id are set by whoever writes the mapper. Two mappers can assign different values to the same vendor event.
  • Version drift. Producers pin a version in metadata.version. Security Lake's AWS sources are on 1.0.0-rc.2 and 1.1.0, while the spec is at 1.9.0. Queries that span versions have to handle classes and attributes that changed in between.

The practical rule is to normalize the sources your detections read, keep the original record for reprocessing, and test each mapper against real events before trusting its output. Deciding which sources earn that work is covered in what security logs to store, and where.

Examples

Here's an illustrative OCSF Authentication event for a failed Okta sign-in, trimmed to the classification fields:

{
  "class_uid": 3002,
  "class_name": "Authentication",
  "category_uid": 3,
  "activity_id": 1,
  "type_uid": 300201,
  "type_name": "Authentication: Logon",
  "status_id": 2,
  "severity_id": 1,
  "user": { "name": "jlee@example.com" },
  "src_endpoint": { "ip": "203.0.113.24" },
  "metadata": { "product": { "name": "Okta", "vendor_name": "Okta" }, "version": "1.9.0" },
  "unmapped": { "debugContext": { "debugData": { "requestUri": "/idp/idx/identify" } } }
}

Once Okta, CloudTrail, and Windows logons are all in that class, one SQL query finds password spraying across all of them. The table name depends on your lake:

SELECT src_endpoint.ip AS source_ip,
       count(DISTINCT "user".name) AS users_tried,
       count(*) AS failures
FROM ocsf_authentication
WHERE class_uid = 3002
  AND activity_id = 1
  AND status_id = 2
GROUP BY src_endpoint.ip
HAVING count(DISTINCT "user".name) >= 10

Onboarding a custom source to OCSF in Cotool shows where the mapping work happens and where to check it:

  1. Connect. Enter the SQS queue URL, the IAM role ARN, and each bucket and prefix Cotool may read. Objects outside those prefixes are skipped even if they reach the queue.
  2. Test. Cotool assumes the role, samples up to ten queue messages, and reads one announced object. Sampled messages go back on the queue.
  3. Format. Choose "Custom format, normalized to OCSF" and read the preview of the sampled events. This is the step to confirm that class_uid, activity_id, and status_id are populated, because a blank enum here becomes a silent miss in every later query.
  4. Enable. Enabling the source requires a passing test.
  5. Export, if a lake needs it. A logging sink copies newly ingested logs to your own bucket as Normalized Parquet or JSON Lines, or as Raw records for reprocessing. Sinks don't backfill data ingested before they were created.

The same idea works across clouds. Cotool's CloudTrail parser unwraps the Records envelope and maps actor, action, service, resource, source IP, outcome, account, and region into columns shared with its GCP Cloud Audit dataset. A detection for an IAM policy change by an unexpected actor then reads one set of columns for AWS and Google Cloud, instead of userIdentity and eventName on one side and GCP's own field names on the other. Digest files are skipped, and each event is keyed by its eventID.

  • Splunk Common Information Model (CIM): Splunk's search-time data models. Detections built on CIM data models are part of what a team rewrites when it evaluates Splunk alternatives.
  • Advanced Security Information Model (ASIM): Microsoft Sentinel's normalization layer of KQL parsers and normalized tables. What it costs to ingest the tables those parsers read is covered in Microsoft Sentinel pricing.
  • Amazon Security Lake: AWS's OCSF-based security data lake, one S3 bucket per Region in your account.
  • AWS Security Finding Format (ASFF): the Security Hub CSPM finding format. Security Hub also emits OCSF findings in Findings Imported V2 events.
  • Modern SIEM: see legacy SIEM vs modern SIEM.

FAQ

What does OCSF stand for?
OCSF stands for Open Cybersecurity Schema Framework. It is an open-source framework for building schemas plus a core schema for security events, published under the Apache License 2.0. The schema browser at schema.ocsf.io lists every OCSF category, event class, object, profile, and extension.
Who created OCSF?
OCSF was conceived and initiated by AWS and Splunk, building on the ICD Schema that Symantec (part of Broadcom) contributed. OCSF launched at Black Hat USA on August 10, 2022 with 15 other initial members, including CrowdStrike, Okta, Palo Alto Networks, and Sumo Logic. OCSF joined the Linux Foundation on November 19, 2024.
What is the difference between OCSF and ECS?
OCSF classifies each event into one numbered event class with required attributes and enum values, and it is independent of any storage engine. ECS, the Elastic Common Schema, defines field names and Elasticsearch data types for storing events in Elasticsearch, and Elastic contributed ECS to OpenTelemetry in 2023. Both normalize security data; OCSF is the one Amazon Security Lake requires.
Does Amazon Security Lake use OCSF?
Amazon Security Lake uses OCSF as its storage schema. It converts natively supported AWS sources to OCSF automatically, and custom sources must be written in OCSF as Apache Parquet. Security Lake's version 2 AWS sources use OCSF 1.1.0, several releases behind the current OCSF 1.9.0.
Is OCSF normalization lossy?
OCSF normalization can be lossy, because fields that don't fit an OCSF attribute go into the optional unmapped object, whose structure consumers shouldn't rely on. Mappers can also keep the original event in raw_data, which the schema says is not meant for structured querying. Keep the raw record somewhere you can reprocess it.

Sources

  1. Open Cybersecurity Schema Framework · accessed November 16, 2026
  2. Understanding the Open Cybersecurity Schema Framework (OCSF docs) · accessed November 16, 2026
  3. OCSF schema API: categories, profiles, and extensions (version 1.9.0) · accessed November 16, 2026
  4. OCSF 1.9.0 Authentication [3002] event class · accessed November 16, 2026
  5. OCSF schema browser · accessed November 16, 2026
  6. ocsf-schema releases (GitHub) · accessed November 16, 2026
  7. Cybersecurity and Technology Industry Leaders Launch Open-Source Project to Help Organizations Detect and Stop Cyberattacks Faster and More Effectively | Splunk · accessed November 16, 2026
  8. AWS co-announces release of the Open Cybersecurity Schema Framework (OCSF) project · accessed November 16, 2026
  9. Open Cybersecurity Schema Framework (OCSF) in Security Lake · accessed November 16, 2026
  10. Convert data in the _raw field to OCSF format | Splunk Edge Processor · accessed November 16, 2026
  11. Data requirements for OCSF conversion | Splunk Ingest Processor · accessed November 16, 2026
  12. What is OCSF (Open Cybersecurity Schema Framework)? | SentinelOne · accessed November 16, 2026
  13. Elastic Common Schema (ECS) reference · accessed November 16, 2026
  14. Elastic contributes Elastic Common Schema (ECS) to OpenTelemetry · accessed November 16, 2026
  15. Open Source Security Events Metadata (OSSEM) README · accessed November 16, 2026
  16. Overview of the Splunk Common Information Model · accessed November 16, 2026
  17. Normalization and the Advanced Security Information Model (ASIM) · accessed November 16, 2026
  18. AWS Management Console sign-in events - AWS CloudTrail · accessed November 16, 2026
  19. System Log query | Okta Developer · accessed November 16, 2026
  20. AWS S3 / SQS log ingestion | Cotool docs · accessed November 16, 2026
  21. Exporting logs | Cotool docs · accessed November 16, 2026
  22. Security Log Storage and Federated Search for Agents - Cotool · accessed November 16, 2026

Keep reading

Attackers are scaling with tokens

Cotool helps defenders operate at machine speed. See how security teams are scaling Detection & Response.

Book a demo