Glossary, 30 terms

The vocabulary of agentic cyber defense

Agentic cyber defense

Software in which AI agents carry out defensive security work end to end: investigating alerts, hunting, containing incidents, and writing and tuning detections, within limits a human sets. It differs from a copilot, which answers questions, and from first-generation AI SOC tools that summarise alerts on top of SOAR. The term is vendor vocabulary; this index uses it because buyers now evaluate the same platforms across seven use cases.

Example: Mate Security names its category Agentic Cyber Defense and treats AI SOC as one capability inside it (mate.security, read 2026-11-03).

AI SOC

A security operations platform in which AI agents triage and investigate alerts that human analysts used to work by hand. The term covers a wide range, from summarisation add-ons to agents that plan their own investigations. Latio Pulse splits the market into a first generation built on SOAR and a second generation that reasons over the organisation's own context.

Example: IDC MarketScape: Worldwide Standalone AI SOC Platforms 2026 named Torq a Leader on 2026-10-20, per Torq.

Security context graph

A store of linked facts about an organisation, such as assets, identities, owners, normal behaviour, tickets and past decisions, that an agent queries during an investigation. Its quality depends on what is stored, where each fact comes from, and how stale facts are retired. Without one, an agent gives every company the same verdict for the same alert.

Example: Mate introduced its Security Context Graph on 2026-02-19; it stores memories with source, confidence and freshness.

Gamebook

Mate Security's term for structured guidance that an agent follows while planning its own investigation and response steps. A Gamebook states the goal, the evidence that matters, the decisions that need a human, and the guardrails, rather than drawing every branch in advance as a SOAR playbook does. Other vendors use workflows or policies for similar control.

Example: Mate launched Gamebooks on 2026-08-26 and quotes one customer moving 711 playbooks to 10 Gamebooks in 3 weeks (grade C).

Playbook

A fixed sequence of automated steps in a SOAR tool, drawn as a flowchart with branches for each expected case. Playbooks are predictable and auditable, and they fail closed when an alert does not match a branch. Large estates often hold hundreds of near-duplicate playbooks, one per tool or alert source.

Example: Mate's replace-SOAR page quotes a Fortune 500 company with 711 SOAR playbooks and an expected 9-month rebuild (read 2026-11-03).

CD/CR

A framework published by Mate Security that pairs continuous detection with continuous response, describing how agents keep detections and response actions current as the environment changes. It is vendor vocabulary, not an industry standard. Treat it as a description of how Mate organises its product rather than a benchmark other vendors are measured against.

Example: Mate published its CD/CR framework on 2026-05-18, per the products list on mate.security.

CPR scoring

Mate Security's method for gating agent actions by Confidence, Precision and Reversibility. The score is computed outside the agent's reasoning loop, so the agent cannot argue its way past a limit. Actions with low reversibility need higher confidence and precision before they move from recommend only to supervised or auto-run.

Example: Mate's platform page describes CPR scoring next to its three autonomy levels (read 2026-11-03).

Autonomy level

How much an agent may do without a human: recommend only, supervised (the agent proposes and a human approves), or auto-run (the agent acts and a human reviews afterwards). Good platforms set the level per action class, not for the whole product. The right level depends on how easily an action is reversed.

Example: Legion Security names Guided mode and Autonomous mode on its home page (read 2026-11-05).

Shadow mode

A deployment stage in which the agent investigates live alerts and records verdicts but takes no action and closes nothing. Analysts keep working as usual, and their decisions are compared with the agent's. Shadow mode is the safest way to measure accuracy on your own alerts before any autonomy is granted.

Example: Mate's platform page shows a first shadow-mode verdict at 09:00 in 'your first 24 hours' (read 2026-11-03).

Auto-closure

The share of alerts an agent closes without a human, usually as false positives. A high rate saves analyst time only if the closed alerts were really benign, so an auto-closure figure needs a false-negative figure next to it. Vendors often state auto-closure without a sample or period.

Example: Mate's home page shows 97 percent false-positive auto-closure; its blog says up to 85 percent (read 2026-11-03).

False-positive rate

The share of alerts that look malicious but turn out to be benign. In detection engineering it is measured per rule; in triage it is the share of the queue an agent should close. A reduction claim needs a baseline: what the rate was before, over what period, on which alert sources.

Example: Exaforce claims a 90 percent reduction in false positives with no baseline or sample (exaforce.com, read 2026-11-03).

True positive

An alert that correctly flags real malicious or policy-breaking activity. Some teams separate benign true positives, where the activity happened but was authorised, such as an admin's approved scan. Labelling alerts as true positive, benign true positive or false positive before a trial is the basis of any accuracy figure.

Example: Legion claims 0 false negatives against senior analyst decisions, meaning no true positives missed, with no sample stated (read 2026-11-05).

Mean time to investigate (MTTI)

The average time from an alert arriving to a verdict on what it is. It is the metric AI SOC platforms most often claim to cut, because investigation is the step agents automate first. A reduction claim needs the baseline period, the alert types and whether queue time is included.

Example: Legion claims an 85 percent MTTI reduction for the average customer within 3 months, with no method (read 2026-11-05).

Mean time to contain (MTTC)

The average time from detection to stopping an incident from spreading, for example by isolating a host or disabling an account. It depends on whether the platform can act in your tools, at what autonomy level, and whether approvals are fast. Few vendors in the index publish it separately from response time.

Example: Exaforce cites Forcepoint at a 14 minute MTTR for P0 incidents, with no baseline (read 2026-11-03).

Mean time to respond (MTTR)

The average time from detection to completed response. Definitions vary: some vendors stop the clock at containment, others at full remediation. Always ask where the clock starts and stops before comparing two MTTR figures from different vendors.

Example: Torq cites an unnamed global pharma customer cutting MTTR by 92 percent, a case-study headline (read 2026-11-03).

Containment action

A step that limits an incident's reach: isolating an endpoint, disabling or forcing a password reset on an account, blocking a sender or domain, revoking a session. Agentic platforms differ on which actions they support in which tools, and on whether each is reversible and logged.

Example: Intezer's Workflows (2026-08-19) can close alerts, isolate hosts and notify analysts on verdicts or schedules.

Reversibility

How easily an automated action can be undone, by whom and how quickly. Revoking a session token is highly reversible; deleting mail items is not. Reversibility should set the autonomy level for each action, and the vendor should say how each supported action is undone.

Example: Mate makes reversibility one of the three factors in its CPR action score (read 2026-11-03).

Evidence trail

The record of what an agent looked at, what it concluded and what it did, kept so a human or auditor can reconstruct a decision later. A useful trail links each action to its trigger, its evidence and its approver. Without one, an automated action cannot pass an audit.

Example: Microsoft logs Security Copilot agent actions in the Purview Unified Audit Log (read 2026-11-03).

Claims audit grade

This index's grade for a published figure. A: measured by a third party with a published method. B: run by the vendor with a sample, a method and a date. C: a number with no method. D: a number that conflicts with the vendor's own other pages. Grades feed the published evidence criterion.

Example: In November 2026, four of 49 figures reached grade B and none reached grade A.

Hyperautomation

Large-scale automation that chains many tools and steps into workflows, usually through a low-code builder and a broad connector catalog. In security it describes the SOAR successors that add AI agents on top of a deterministic workflow engine, keeping predictable steps predictable while agents handle judgement.

Example: Torq describes its AI SOC platform as built on a hyperautomation engine (torq.io, read 2026-11-03).

Detection-as-code

Managing detection rules like software: stored in a repository, versioned, reviewed through pull requests and tested before release. It makes agent-proposed rule changes reviewable, because each change is a diff a human can read and a test suite can check.

Example: Prophet says learned changes are previewed and backtested before they apply, the same discipline (read 2026-11-03).

Rule health

Whether a detection rule still works: its log source still arrives, its fields still exist, and it fires when it should. Rules often break silently after a format change. Monitoring rule health catches gaps that look like quiet days.

Example: Mate lists health with tuning and coverage under Agentic Detection Engineering (mate.security, read 2026-11-03).

Coverage gap

An attacker technique relevant to your organisation that no detection would catch. Gaps are usually measured against MITRE ATT&CK. The denominator matters: coverage against all techniques flatters nobody, so pick the techniques from threat reports relevant to your sector.

Example: Simbian says the best of 32 products covers 45.1 percent of ATT&CK tactics; the runner is not stated (read 2026-11-03).

ATT&CK technique

An entry in MITRE ATT&CK, the public catalogue of attacker behaviour grouped by tactic, such as initial access, persistence or exfiltration. Techniques give hunting and detection engineering a shared vocabulary and let teams measure coverage technique by technique.

Example: Simbian reports that only 7 of 13 products it lists clear 50 percent of ATT&CK tactics (simbian.ai, read 2026-11-03).

Insider risk

The risk that someone with legitimate access, an employee, contractor or partner, misuses it, by intent or by mistake. It overlaps with data loss prevention (DLP). Agentic platforms address it by watching data movement and behaviour against what is normal for that person and role.

Example: Microsoft Security Copilot's Purview agents cover data security; Legion lists DLP among its use cases (read 2026-11-03 and 2026-11-05).

Exposure management

Finding and prioritising what an attacker could reach and exploit: vulnerable systems, misconfigurations, over-privileged identities. It moves the SOC from reacting to alerts toward closing paths before they are used. Few agentic platforms document it beyond a mention.

Example: Mate lists exposure management as a one-line tile with no product page (mate.security, read 2026-11-03).

MDR

Managed detection and response: a service in which an outside team watches your environment and responds to threats, usually around the clock. Several agentic platforms now sell an MDR-style service on top of their agents, mixing human analysts and AI.

Example: Exaforce offers a managed 24/7 SOC run by its analysts and Exabots (exaforce.com, read 2026-11-03).

MSSP

Managed security service provider: a company that runs security operations for many clients. An MSSP needs a platform that works across many client tool stacks and separates tenants cleanly, so integration breadth and pricing per client matter more than for a single enterprise.

Example: Mate names Water-IT as an MSSP running Mate across client stacks (mate.security, read 2026-11-03).

SOAR

Security orchestration, automation and response: tools that run playbooks across security products to automate repetitive steps. SOAR made automation possible but brittle, because every case needs a drawn branch. Agentic platforms either replace SOAR or add agents above a SOAR-like engine.

Example: CrowdStrike launched Charlotte Agentic SOAR on 2025-11-04, combining agents with deterministic workflows.

SIEM

Security information and event management: the system that collects logs and alerts, stores them and runs detection rules over them. Most agentic defense platforms read from a SIEM rather than replacing it, though some now sell their own data layer.

Example: 7AI announced a Federated SIEM on 2026-07-27 (7ai.com, read 2026-11-03).