← all systems · see it in The System
Security orchestration, automation and response (SOAR)
A group of cybersecurity technologies that allow organizations to respond to some incidents automatically: it collects inputs monitored by the security operations team (alerts from the SIEM, threat intelligence platforms and other tools) and helps define, prioritize and drive standardized incident response activities (src:wp:security-orchestration). Researchers define security orchestration as 'the planning, integration, cooperation, and coordination of the activities of security tools and experts to produce and automate required actions in response to any security incident across multiple technology paradigms' (src:islam-2019-mlr). Walked as the platform and the practice: integrations (connectors, APIs, plugins), data processing and enrichment, case and ticket management with a knowledge base, the playbook engine that orchestrates tools and people, and the analysts who write the playbooks and approve what is not automated.
Properties published scienceour proposal
Sourced fields, and our own readings.
| aliases | SOAR and automation |
| born at | 2015 or 2017 (term; C-3983); security orchestration literature from 2007 (src:islam-2019-mlr) and response-automation research from the early 2000s (src:huang-2025-intent-soar) as forerunners |
| born precision | years |
| boundary | Inside: the integration layer (connectors, wrappers, plugins, API gateway, tool registry), data processing and enrichment, the semantic layer and knowledge base, ticket and case management, the orchestrator and task manager that run playbooks (workflows, runbooks), the threat-intelligence management engine, dashboards and the user interface, and the playbooks themselves. Outside: the security tools it drives and reads (SIEM, IDS/IPS, EDR, firewalls, threat-intelligence platforms, ticketing systems, which other walks own), the organisation's estate, the attackers, and the SOC and incident responders who write and approve. Whether the analysts are part of SOAR or users of it is contested (C-3982): the review literature calls the human 'an essential part of a security orchestration platform'; the encyclopedia calls SOAR 'a group of cybersecurity technologies'. |
| nested level | A technological system (rung 12): software running inside a security operations centre (its response and coordination floors), inside an organisation; built on the APIs of the security tools it connects; its top floor is run by people. |
| scale | technological |
Measures (29) published science
| alerts reaching the average SOC (Dark Reading, cited) | 10,000 alerts per day |
| false-positive rate, upper figures cited | as high as 80 percent |
| survey of 500 decision-makers on alerts | 35% missed an alert; 22% ignored alerts entirely; 25% did not act on a high-profile alert percent of respondents |
| true positive alert ratio in one dataset (cited) | 1.59 percent |
| incident timeline that motivated the field: breach to discovery, then to remediation (BakerHostetler, cited) | 61 days to discover; 41 more days to remediate |
| alerts actively investigated (report cited) | less than 10 percent |
| market forecast | 1.6 billion by 2021 (forecast) USD |
| Navy SOAR user study size | 24 participants; 6 commercial tools; over 200 hours people; tools; hours |
| customization per tool at installation | 2-4 hours per tool |
| tools completing a playbook with internet denied | 1 of 6 tools |
| ticket accuracy against the manual baseline | all six tools worse than the baseline comparison |
| APIs to accommodate per tool when integrating (cited) | over 200 APIs per tool |
| analysts' comfort with AI agents, response phase against identification phase | 2.71 against 4.04 (p < 0.001) mean rating |
| share of incident response operations automatable to some extent (vendor claim, cited) | 80-90 percent |
| work automated by one vendor (vendor claim, cited) | 800,000 man-hours in two years; $38.5 million hours; USD |
| attacker engagement time, static against SOAR-placed honeypots | 102 against 3148 seconds per instance |
| CACAO playbooks generated from CSAF advisories | 71 of 79; precision 98.72%; recall 78.57% playbooks; percent |
| AI-for-SOC papers focused on detection | 65 (of 189 papers); 88% non-explainable approaches percent |
| honeypot orchestration over four days of live traffic | 40 orchestrations; 7823 attacks; 965 DDoS packets; 3 malicious samples counts over 4 days |
| bandwidth while monitoring at rest | 0.373 ± 0.621 MB per 1,000 events |
| bandwidth to work one investigation | 22.80 ± 15.73 MB |
| LLM use in one SOC over 10 months | 3,090 queries from 45 analysts queries; analysts |
| API recommendation for building integrations (APIRO) | 91.9 percent Top-1 accuracy |
| Gartner forecast of adoption by 2019 (cited) | 30 percent of large and medium enterprises (forecast) |
| malware triage time in the Navy test range, unassisted against SOAR-assisted | about 15 minutes against almost instantaneous |
| vendor configuration window in the Navy test | 13 business days; hands-on only the final 7 |
| 'insufficient automation level of SOC components' raised, analysts against managers | 33.33 against 77.78 percent of participants |
| mitigation time targets by priority, and success rate, in one internal SOC (P11) | P0 30 min, P1 60 min, P2 120 min, P3 24 h, P4 48 h; 87% met minutes; hours; percent |
| the SOC clocks a SOAR's cases feed (named, values site-specific) | mean time to detect, react, triage, escalate, contain and resolve; number of incidents handled automatically named clocks |
Open contentions (5) open contention
Where sources or readings disagree. Both stay on the map.
- C-3981 Will security-automation platforms stay a product of their own or be absorbed into bigger tools?
booked: The map keeps the competing readings; none is ruled.
the other reading: unclear whether it will become as ubiquitous as other security tools; AIOps may be bolted on (src:horneman-ray-2021-sei); SOAR features join SIEMs (SIEM walk) and LLM agents work through SOAR at a semi-autonomous level (src:srinivas-2025-ai-soc); trade press in 2026 reports a vendor's 'SOAR is dead' and a Gartner analyst's view that it will morph into agentic AI SOC software (src:nadeau-2026-techtarget, read only through a summarising fetch) - C-3983 When was security orchestration and automation named as a category: 2015 or 2017?
booked: The map keeps the competing readings; none is ruled.
the other reading: 2015: "2015, when Gartner coined the SOAR term" (src:horneman-ray-2021-sei); 2017: "the term 'SOAR' was coined by Gartner only in 2017" (src:bridges-2023-soar-test) - C-3984 How brittle are automated security playbooks in practice?
booked: The map keeps the competing readings; none is ruled.
the other reading: playbooks are hard-coded for known alerts and rigid, useless against novel ones (src:kremer-2023-ic-secure); playbooks enforce low-level configuration without the analysts' intent; intents proposed instead (src:huang-2025-intent-soar); playbook upkeep takes considerable time (participant P5, src:bridges-2023-soar-test); playbooks are standardised and generated from structured advisories with high precision (98.72%) (src:empl-2023-ics-playbooks); the upkeep loop can end in disuse: misconfigured or outdated SOC tools can lead an organisation to abandon them (src:bridges-2023-soar-test citing Sundaramurthy et al.) - C-3985 Does automating a security team's work make it better, or only faster?
booked: The map keeps the competing readings; none is ruled.
the other reading: faster: all six tools as fast or faster than the baseline, more so when well configured (src:bridges-2023-soar-test); worse records: all six tools wrote less accurate tickets than manual work; analysts' own reports more complete (src:bridges-2023-soar-test, src:tilbury-2024-automation-bias); 'some small reduction of alert fatigue' at the price of complexity (src:horneman-ray-2021-sei); vendor claims of large savings (800,000 hours; src:islam-2019-mlr citing Hexadite); scope: the controlled study's scenarios were investigation and documentation in a test range, vendors had 13 business days (7 hands-on), participants learned each tool from a vendor tutorial; in that range automated malware triage went from about 15 minutes to almost instantaneous (src:bridges-2023-soar-test). Automated containment in deployed operations is not measured in any open source reached - C-3986 Which security actions should never run without a person approving them?
booked: The map keeps the competing readings; none is ruled.
the other reading: containment of a breach and unknown threats need a human in the loop; notifying, assigning and enrichment are safe to automate (src:islam-2019-mlr, citing practitioners); higher levels of automation suit well-defined processes and well-modelled, less severe threats; ambiguous processes stay with analysts (src:tilbury-2024-humans-automation); 80-90% of incident response operations automatable to some extent (vendor claim, src:islam-2019-mlr citing Swimlane); autonomous agents and intent-based SOAR take over response (src:huang-2025-intent-soar, src:shaikh-2026-red-team-soar); analysts want automation before use and worry about too much after (src:bridges-2023-soar-test); comfort with AI agents lower for response than identification (src:tilbury-2024-humans-automation); a standard automated step (disconnect the device) can itself cause harm where the device does critical work (src:shaked-2023-playbooks); the disagreement also runs along roles: 'insufficient automation' raised by 77.78% of managers against 33.33% of analysts; seven managers 'always want more automation'; an analyst: management 'not realizing that they are getting less quality' (src:kokulu-2019-socs); automated blocking can be abused by attackers to deny access to legitimate users; NIST advises extreme care (src:nist-sp800-31)
Cross-scale patterns seen here (16) our proposal
Shapes that recur when the same kind of problem is solved at different scales, and how each shows in Security orchestration, automation and response (SOAR). Each links to every other system that shows it.
- same role, different builds (structural)
One job, carrying out a response written in advance for a recognised kind of event, faster than deliberation, with a person kept for the uncertain cases: SOAR playbooks run on SIEM alerts; clinical decision support gives clinicians pre-built advice and alerts; the spinal withdrawal reflex acts below - templating (structural)
A playbook written once in an open format (CACAO) is generated from vendor advisories (71 of 79) and can be read by different platforms; the semantic layer unifies playbooks from different vendors so one plan runs on another tool set. - trade one property for another (structural)
Speed bought with completeness: every SOAR tool tested was as fast or faster than the manual baseline, and every one wrote less accurate tickets; the six tools spread along the trade (Tool 2 fastest with fewest window switches, Tool 5 preferred by users but worst on tickets). A second trade: higher - Umwelt: each system carves its own slice of one shared medium (structural)
Its slice of the estate is exactly what its integrations expose: alerts the detectors raise, answers to the queries its playbooks ask, and what its decoys record. Everything else in the same network is outside its world. - who opens the door (structural)
A port opened ahead of time by the receiver: a SOAR reaches a tool only through a connector its installers built (two to four hours of customization per tool); a tool without one stays outside ("each technology that is unable to be integrated hampers the benefits"). - borrowed-then-held (partial)
Response steps first done by analysts by hand (the borrowed phase: an expert goes to the endpoint tool, then commands the firewall) are later held in playbooks the platform runs; automation was described as "mimicking ideal steps of a human". Partial: the people who did the steps remain inside or be - holding / entrainment (partial)
The engine sets the number and the death of its decoy parts: it deployed and re-deployed honeypots 40 times in four days according to the attacker's behaviour. Partial: one research engine. - partner-becomes-part (partial)
Separate products (orchestration and automation tools, incident-response platforms, threat-intelligence platforms) became one platform category; the SIEM walk records SOAR in turn joining SIEMs as a native feature. Partial: fusion of products, inheritance not tested. - red-queen (partial)
Both sides automate: SOAR platforms "must rapidly adapt to continuously evolving cyber attacks", and researchers build autonomous LLM-and-RL attackers to test autonomous SOAR defenders. Partial: both sides improve; the stalemate (neither gaining ground) is not shown in the sources read. - same shape, opposite sign (partial)
One process, a two-tier automated planner (a strategic layer stating intent, a tactical layer choosing and running tool actions), built for defence (intent-based SOAR: security intents over playbook actions) and for attack (an LLM planner for strategic intent with an RL controller for tactical execu - same toolkit, independent builds (partial)
Six commercial tools built from the same defining parts (ingest, correlate and prioritise, automate by playbooks, collaborate) varied greatly in strengths and weaknesses. Partial: whether the builds are independent or share lineage is not stated. - The carry (partial)
Two faces. The wait: steps that need a person wait for approval (levels 4-5: automation selects, the analyst approves) and the chain runs at the analyst's pace. The lookahead: enrichment gathers the context before the analyst opens the case, so the analyst does not wait for each lookup. Overflow is - The Payment to Entropy (partial)
Signal form, at the hand-off from investigation to record: auto-populated tickets lost fields that operators wrote by hand ("some fields were populated for some investigations but not for others"); SOAR-generated alert reports were less complete than analysts' own. Partial: the loss is measured as c - nested ladder (surface)
Levels of automation as a ladder: L3 offers options, L4 selects and asks approval, L5 implements if approved, L6 implements unless reversed in time, L7 acts and informs, L8 informs only if asked; each level hands more of the decision to the machine. Surface: the levels are a design scale the review - runaway from an unstable balance (surface)
More tools raise more alerts, which push SOCs to add more automation, which raises reliance and alerts again (fb-automation-bias); SOCs "rush toward implementing increased automation". Surface: no balance point is shown. - the whole sets the local reference (surface)
The organisation's policy sets, for every playbook step, whether it may run alone or needs a person; each step reads its own rule. Surface: no case where local readings fail to close is shown.
Valence: what it trades, and with whom (13) published science
Each exchange with a neighbour, read from both sides: who gains, who pays, and what carries it.
- Human (Homo sapiens) via sys:electromagnetic-field · Security orchestration, automation and response (SOAR) gain
- Large language model via sys:written-record · Security orchestration, automation and response (SOAR) gain
- Intrusion detection system via sys:written-record · Security orchestration, automation and response (SOAR) gain
- Security operations center (SOC) via sys:written-record · Security orchestration, automation and response (SOAR) gain
- Incident response via sys:written-record · Security orchestration, automation and response (SOAR) gain
- Threat actor via sys:computer-network · Security orchestration, automation and response (SOAR) gain
- Organization via sys:written-record · Security orchestration, automation and response (SOAR) gain
- Firewall (computing) via sys:written-record · Security orchestration, automation and response (SOAR) gain
- Security information and event management (SIEM) via sys:written-record · Security orchestration, automation and response (SOAR) gain
- Honeypots (decoys) via sys:computer-network · Security orchestration, automation and response (SOAR) gain
- Cyber threat intelligence via sys:written-record · Security orchestration, automation and response (SOAR) gain
- CACAO security playbook (OASIS open standard) via sys:written-record · Security orchestration, automation and response (SOAR) gain
- IT service management and ticketing (Jira, ServiceNow) via sys:written-record · Security orchestration, automation and response (SOAR) gain
Partners and parts published science
Built from / builds (11)
- Security operations center (SOC)
- The attacker's time: real targets against SOAR-placed decoys
- Platforms competing for the SOC's response-automation role
- Playbooks and standard operating procedures
- Threat intelligence platform (TIP)
- Data processing and enrichment (data curator, extractor, analyzer)
- Semantic layer and knowledge base (ontology of tools, capabilities, IRP activities)
- Ticket and case management (SIRP functionality)
- Orchestrator and task manager (playbook engine)
- Dashboard and user interface
- Integrations: connectors, wrappers, plugins, API gateway, tool registry
Sits inside / holds (3)
expressed by (25)
- Generation/Synthesis
- Scaffolding/Enablement (Meta-Tooling)
- Templating / Standardization
- Quality Improvement/Refinement
- Interoperability/Transduction (Bridging)
- Encapsulation & Modularity (Abstraction)
- Symbiosis/Co-Creation (Collaborative Specialization)
- Internal Coordination/Regulation
- Sensing/Scanning
- Mapping/Modeling
- Uncertainty Reduction/Learning
- Multi-Scale Data Fusion / Triangulation
- Trail-Marking / Externalized Memory (Stigmergy)
- Frontier Outposts / Sensing Extension
- Social Learning / Public Information Use
- Speed
- Quick Harm Avoidance
- Reactive Response
- Proactive Defense
- Self-Monitoring
- Validation of Effectiveness
- Deception
- Resilience & Repair
- Redundancy & Diversity
- Triage & Prioritized Allocation
floor filled by (17)
- Human (Homo sapiens)
- Large language model
- Intrusion detection system
- Security operations center (SOC)
- Threat actor
- Organization
- Firewall (computing)
- Security information and event management (SIEM)
- Honeypots (decoys)
- Playbooks and standard operating procedures
- Threat intelligence platform (TIP)
- Data processing and enrichment (data curator, extractor, analyzer)
- Semantic layer and knowledge base (ontology of tools, capabilities, IRP activities)
- Ticket and case management (SIRP functionality)
- Orchestrator and task manager (playbook engine)
- IT service management and ticketing (Jira, ServiceNow)
- Integrations: connectors, wrappers, plugins, API gateway, tool registry
Kinds (10)
- Threat intelligence platform (TIP)
- Ticket and case management (SIRP functionality)
- Security orchestration and automation (SOA)
- Centralized SOAR deployment
- Distributed SOAR deployment
- Cloud-based SOAR tool
- AI-enabled (autonomous-agent) SOAR
- Intent-based SOAR (autonomic cyber defence)
- Security orchestration in software-defined and virtualised networks (SDN/NFV)
- Security analytics
Notes published science
- Before they tried them, Navy SOC analysts named automation as the feature they most wanted in a SOAR tool; after working investigations with six commercial tools in a test range, they worried about too much of it. Every tool was as fast as doing it by hand or faster, and every tool wrote worse tickets than the analysts did by following the procedure step by step (src:bridges-2023-soar-test).
- A SOAR engine that moves its honeypots to wherever an intruder is poking around kept attackers busy on fake machines for about 3,148 seconds on average, against 102 seconds for honeypots that stayed put: roughly 30 times longer (src:bartwal-2022-soar-honeypots).
- Analysts are comfortable letting AI help them spot an attack (4.04 on the rating scale) but much less comfortable letting it respond (2.71): the closer the machine gets to acting on the world, the more the people want their hand on it (src:tilbury-2024-humans-automation).
- The tool that runs a security team's playbooks is also the tool that keeps the team's scorecard: one SOC's SOAR produced most of its performance numbers, and because it can auto-close most false alarms, a low false-alarm rate can look like skill when it is really automation; analysts elsewhere called such numbers 'a tool that their managers use to show false improvement' (src:forsberg-frantti-2023, src:kokulu-2019-socs).
Layers
The ways it takes in its world, where they meet, and how it acts back, each from L1 (where it meets the world) up.
SOAR as security operations engineers and analysts describe it: the estate's alerts and the tools that raise them, up to the response decisions it carries out (reader's view)
- L1: The organisation's estate and what happens in it: hosts, networks, users and the attacks on them, as alerts and records
- L2: The security tools it drives and reads: SIEM, IDS/IPS, EDR, firewalls, threat-intelligence platforms, ticketing and team communication
- L3: Integration: connectors, wrappers, plugins, API gateway and tool registry
- L4: Data processing: alerts filtered, parsed and enriched with context and threat intelligence
- L5: Cases and knowledge: ticket and case management, the knowledge base of tools, capabilities and plans
- L6: Orchestration: the playbook engine (orchestrator and task manager) chooses and runs the workflow across tools
- L7: The analysts and the organisation's policy: they write the playbooks, approve what is not automated, and decide; the user interface is where they meet it
the boardroom (shared floors)
- L5: Cases and knowledge: ticket and case management (the incident-response platform function), the knowledge base of tools, capabilities and plans, and the record of past incidents
- L6: Orchestration: the playbook engine maps an alert to a playbook (by rules the analysts wrote), runs its steps across tools, and holds or forwards steps that need approval
- L7: The analysts and the organisation's policy: what is automated and to what level, which actions need a person, playbook design and approval, the decision on unknown threats
way in: Alerts from the SIEM and other detectors
- L1: Activity in the estate that a detector flags
- L2: The detectors: the SIEM, IDS/IPS, endpoint tools
- L3: Ingested through the integration (connector or API) for that tool
- L4: Filtered for syntactic and semantic correctness, then parsed
way in: Analysts' commands, approvals and overrides
- L1: The analyst's judgement, entered by hand at the console
- L2: The user interface and its APIs
way in: Answers to its own queries (enrichment and investigation results from tools)
- L1: The state of hosts, addresses and accounts in the estate
- L2: The tools queried: endpoint defence, scanners, reputation services
- L3: Queries and replies through the API gateway and wrappers
- L4: Context extracted and generated around the alert
way in: Attacker behaviour on its decoys (honeypot telemetry)
- L1: An attacker's moves inside the internal network
- L2: Honeypots it deployed record the activity, botnet traffic and DDoS
way in: Its own state and metrics on the dashboard
- L1: Its scenarios, assets and runs, counted
- L2: Shown on the dashboard
- L5: Case dispositions (false positive, true positive, benign true positive) and timestamps: the raw data of the SOC's measures
way in: Model predictions (ML trained on past incidents)
- L1: The record of past incidents
way in: Playbooks and standard procedures written for it
- L1: The organisation's incident response plans, vendors' advisories and shared playbooks
- L2: Written as text, flow diagrams, tables, or machine-readable markup (CACAO)
- L3: Generated or imported from structured advisories (CSAF to CACAO)
- L4: Interpreted against the knowledge base so playbooks from different vendors share one meaning
- L5: Stored in the knowledge base as IRP activities
way in: Results of the actions it ran (success, stall, error)
- L1: The estate after an action: blocked, isolated, or unchanged
- L2: The tool that executed the command answers (or does not)
- L3: Reply or error returned through the integration
way in: Status of the connected tools (the tool registry)
- L1: The tools available in the organisation and their capabilities
- L3: The tool registry discovers, registers and watches them
way in: The organisation's policy and rules on what may run without a person
- L1: Organisational policies and rules
way in: Threat intelligence feeds
- L1: Other organisations' and researchers' observations of threats, published as indicators
- L2: Feeds and threat-intelligence platforms
- L3: Taken in through feed integrations
- L4: Aggregated and deduplicated, then matched against alerts
way out: Containment: block addresses, isolate hosts, change firewall rules
- L1: Traffic blocked or a device cut off on the network
- L2: The firewall or endpoint tool carries it out
- L3: Sent through the integration to the tool's API
- L6: A playbook step (automatic or approved) commands the action
way out: Decoys deployed: honeypots placed where the attacker moves
- L1: A false target appears on the network
- L6: The engine decides to deploy a honeypot based on the attacker's behaviour
way out: Notify and assign: alerts, approval requests and assignments to analysts and stakeholders
- L1: Light from screens to the analyst's eyes
- L2: Shown on the dashboard or sent through team communication tools
- L6: Notification, assignment or a request for approval
way out: Removal: a phishing email or malware taken out, each by its own tool
- L1: The message or file removed from the mail store or host
- L6: An IRP step 'remove phishing email' or 'remove malware'
way out: Threat intelligence pushed downstream to the SIEM and firewall
- L1: Indicators installed in the receiving tools
- L6: Aggregated intelligence distributed
way out: Tickets and case documentation
- L1: The written ticket (in the tool or in Jira-like ticketing)
- L5: Case fields auto-populated from the investigation