Security conversations are full of acronyms, and AI has added a whole new layer. Here are the identity and AI security terms we hear most often, explained in plain English. Terms marked “Cygrid” are specific to the Cygrid platform.

A

Access review. A periodic check of who has access to what, so that unneeded permissions can be removed. Often required by auditors and regulators.

Agentic AI. AI systems that do not just answer questions but take actions on their own, such as sending emails, updating records or calling other tools.

AI agent. A piece of software, usually powered by a large language model, that acts on someone's behalf inside business systems. From a security point of view, every agent acts through an AI identity with its own access.

AI agent governance state (Cygrid). How far your organization controls an AI agent: shadow (not known or approved), approved (allowed, but its actions aren't checked against policy) or enforced (its actions go through the Cygrid MCP Gateway and are checked against your policies). For example, an unreviewed Claude connector is shadow, a Copilot agent cleared by security is approved, and a Claude agent routed through the MCP Gateway is enforced.

AI Gateway (Cygrid). Cygrid's own enforcement point in front of LLM providers. Each prompt and model call from agents or applications is tied to an identity, decided by Cygrid against policy and the Identity Intelligence Graph, and allowed, blocked or constrained in real time. It also covers agents that don't use MCP.

AI identity. An identity used by an AI agent or AI tool to act on systems on behalf of a user or organization. For example, a Claude connector authorized in Google Workspace, a Copilot Studio agent acting on SharePoint, or Cursor authorized as a GitHub OAuth app.

API token. A credential issued by a platform for programmatic access, not tied to a person or to a separate account. The token itself is the identity, and its access is defined by the scopes or permissions attached to it. For example, a HubSpot private app access token or a Shopify Admin API access token.

App Enforcer (Cygrid). Enforces Cygrid's decisions out of band, directly in SaaS apps through their APIs. It runs as one-time cleanups, such as removing stale or over-privileged tokens, and continuously on events received via webhooks: it revokes or restricts tokens and permissions and triggers workflows when the Identity Intelligence Graph flags a risk. It never sits in the request path.

Audit trail. A record of who did what, when and why. A good audit trail links every action, including actions taken by AI agents, back to a responsible person.

B

Bot. A non-human identity with a user-like presence in a platform: it appears in member lists and can be messaged or mentioned, but it acts programmatically, not as a person. For example, a Slack bot that posts deployment alerts to a channel.

C

Connected app. A third-party or internal app granted OAuth access to a platform, so it can read or act on data on behalf of users or the organization. Its access is defined by the scopes it was granted. For example, an analytics tool with read access to Salesforce records, or the Jira app installed in a Slack workspace.

D

Decision log. A tamper-proof record of every access decision: what was requested, by whom and on whose behalf, which policy applied and what the verdict was. In Cygrid, every verdict is written to an immutable decision log.

Detection. Suspicious or anomalous activity by an identity, flagged when it matches a detection rule. It describes something that has happened, not a risky state. For example, a HubSpot user exporting all contacts, or a service account calling APIs it has never used before.

DORA. The EU Digital Operational Resilience Act. It sets requirements for how financial organizations manage ICT risk, including access control and third-party risk.

Downscoping. Reducing what an identity or a single action is allowed to do, for example letting an agent export 100 records instead of 100,000, rather than blocking it completely.

E

Enforcement points (Cygrid). The places where Cygrid's decisions are enforced, in two bands: after grant (App Enforcer) and at runtime (MCP Gateway, AI Gateway and Gateway Connect). Every decision is made by Cygrid, against your policies and the Identity Intelligence Graph.

Entitlement. The access an identity has in a system, made up of the roles, permissions and scopes assigned to it. For example, a HubSpot Super Admin role, a Google gmail.send scope or admin permission on a GitHub repository.

EU AI Act. The European Union's law on artificial intelligence. It sets obligations based on risk level, including transparency, human oversight and governance.

External user. A human identity belonging to someone outside your organization, such as a contractor, consultant or agency, who has a full account in your systems in order to work with you. For example, a contractor's account in your identity provider with access to GitHub and Jira.

F

Finding. A specific instance of a risk in your current configuration, such as one identity with excessive access. It describes a risky state that could be exploited, not an event that has already happened. For example, a HubSpot Super Admin who hasn't signed in for 90 days, or a GitHub personal access token with no expiry.

G

Gateway Connect (Cygrid). Turns a gateway you already run, such as LiteLLM or Kong, into a Cygrid enforcement point. Before each agent or LLM request goes through, the gateway asks Cygrid, which evaluates identity, intent and context against policy and the Identity Intelligence Graph and returns allow or deny. Your gateway enforces, Cygrid decides, with no new proxy to deploy.

Guest. A lightweight human identity created so someone outside your organization can collaborate in specific spaces, such as a channel, shared folder or project, without a full account. For example, a Microsoft Teams guest added to a project channel, or a single-channel guest in Slack.

H

Human identity. An identity operated by people or tied directly to a person, including credentials that act with a person's own permissions. For example, an employee's Entra ID account, a shared helpdesk@ account or a developer's GitHub personal access token.

I

Identity and access management (IAM). The policies and tools that manage digital identities and control their access to systems and data.

Identity Intelligence Graph (Cygrid). Cygrid's single graph of humans, non-human identities, AI agents, tokens, MCP servers and resources. It shows what each identity may do, what it actually does and on whose behalf, so risk can be judged in context.

Identity provider (IdP). The system that stores identities and handles sign-in, such as Microsoft Entra ID or Google Workspace.

Identity threat detection and response (ITDR). Tools and practices for spotting attacks that target identities, such as account takeover or abuse of privileges, and responding to them.

J

Just-in-time (JIT) access. Granting elevated access only when it is needed and only for a short time, instead of leaving it permanently in place.

L

Least privilege. The principle that every identity should have only the access it needs to do its job, and nothing more.

M

MCP Gateway (Cygrid). Cygrid's own enforcement point between AI agents and the MCP servers they connect to. Each tool call is tied to an identity, decided by Cygrid against policy and the Identity Intelligence Graph, and allowed, blocked or constrained in real time, with a full audit trail of what each agent did. For example, an agent can be allowed to read Jira issues but blocked from creating them.

MCP server. A service that makes a platform's data and actions available to AI agents through MCP. For example, Atlassian's MCP server or HubSpot's MCP server.

MCP tool. A single action an MCP server makes available to an AI agent, such as creating a Jira issue, searching HubSpot CRM records or deleting a ClickUp task. The Cygrid MCP Gateway applies your policies to each tool call.

Model Context Protocol (MCP). An open standard for connecting AI agents to external systems, so they can read data and take actions in them. For example, a Claude agent connecting to Jira, or a Cursor agent connecting to GitHub.

Multi-factor authentication (MFA). A sign-in that requires two or more independent proofs of identity, such as a password plus a code from an authenticator app, a security key or a passkey. MFA protects the sign-in itself, not the session cookies or tokens issued after it, so a stolen token can still be used without a second factor.

N

NIS2. The EU directive on network and information security. It extends cybersecurity obligations to many more sectors and makes management accountable for them.

Non-human identity (NHI). An identity that isn't operated by a person or tied to one, used by applications, automation or shared resources. For example, a Google Workspace service account, a Salesforce connected app, a HubSpot private app token, a Slack bot or an Entra meeting-room account. AI agents are tracked separately as AI identities.

O

OAuth. An open standard that lets an app access a platform on a user's or organization's behalf without receiving their password. The app gets a token limited to the scopes it was granted. For example, a user granting a scheduling app access to their Google Calendar on a consent screen, or an admin installing a Slack app for the whole workspace.

Orphaned account. An account that no longer has an active owner, for example after someone leaves the company. Orphaned accounts are easy for attackers to misuse.

Owner. The person accountable for an identity. Every identity, whether human, non-human or AI, should have an owner, and an owner is always a real person, never a team or another account. Identities with no identified owner are marked as having an unknown owner.

P

Personal access token (PAT). A token a person creates to access a platform's API or command-line tools as themselves. It acts with that person's permissions, sometimes narrowed by scopes. For example, a GitHub token a developer uses to clone repositories from the command line.

Privileged access management (PAM). Controls for the most powerful accounts, such as administrators, including approval, monitoring and session recording.

Prompt injection. An attack in which malicious instructions, typed directly or hidden in documents, web pages or other content the model reads, trick an AI model into doing something it should not.

R

Risk level (Cygrid). A rating of how serious a finding or detection is, based on its potential impact and how likely it is to be abused. Levels are Info, Low, Medium, High and Critical.

Runtime enforcement. Checking and deciding on each action at the moment it happens, rather than relying only on permissions set in advance. In Cygrid, the MCP Gateway, AI Gateway and Gateway Connect enforce at runtime.

S

Scope. A permission granted to an app or token that defines which data it can access and which actions it can perform. When the app acts on a user's behalf, it can't exceed that user's own permissions. For example, Google's gmail.send, HubSpot's crm.objects.contacts.write or Slack's channels:read.

Scope sensitivity (Cygrid). A rating of how much access a scope grants, based on how sensitive the data or actions behind it are. Levels are Non-Sensitive, Sensitive and Highly Sensitive. For example, viewing a user's basic profile is Non-Sensitive, reading calendar events is Sensitive and reading all Gmail messages is Highly Sensitive.

Service account. A user-like account created for automation or an integration, not for a person. It holds its own permissions on the platform, rather than borrowing them from a user. For example, a Google Cloud service account or a Salesforce integration user.

Shadow AI. AI tools and agents that employees use or connect without approval from IT or security. In Cygrid, these agents have the governance state shadow.

Shared account. A human identity used by more than one person. Because several people sign in with the same credentials, actions can't be traced to one individual. For example, a helpdesk@ account used by several support engineers.

Z

Zero trust. A security model that never assumes trust based on network location and verifies every user, device and request.

Missing a term? Let us know and we will add it. And if you want to see how these ideas work together in practice, take a look at the Cygrid platform.

Masha Kovaleva

PR & Marketing Manager

Share

Stay in the loop

Product news and identity security insights.

Platform

Solutions by Team

Solutions by Industry

Company

Legal

© 2026 Cygrid GmbH. All rights reserved.

Stay in the loop

Product news and identity security insights.

Platform

Solutions by Team

Solutions by Industry

Company

Legal

© 2026 Cygrid GmbH. All rights reserved.