Glossary

Attribute-Based Access Control

ABAC (Attribute-Based Access Control) decides access by evaluating policies over attributes of the user, resource, action, and environment, expressing fine-grained contextual rules that roles alone cannot.

FIDO2 / WebAuthn

FIDO2 is the FIDO Alliance's open-standard set for phishing-resistant, passwordless authentication with public-key cryptography, comprising the WebAuthn browser API and the CTAP authenticator protocol.

JSON Web Token

A JWT (JSON Web Token, RFC 7519) is a compact, URL-safe token of signed JSON claims that a recipient can verify without a database lookup; OIDC ID tokens are JWTs.

Multi-Factor Authentication · 2FA · Two-Factor Authentication

MFA (Multi-Factor Authentication) requires two or more independent factors (something you know, have, or are) to prove identity, so a stolen password alone cannot log an attacker in.

OAuth · OAuth2

OAuth 2.0 is a delegated authorization framework (RFC 6749) that lets a user grant an app scoped, time-limited access to their data on another service without sharing a password; it is authorization, not authentication.

OpenID Connect

OIDC (OpenID Connect) is a thin authentication layer built on OAuth 2.0 that adds an ID token, a signed JWT the client reads to verify who the user is; it powers most modern consumer login.

Passkeys

A passkey is a discoverable WebAuthn credential, usually synced across a user's devices, that replaces a password with a phishing-resistant key pair held by the platform.

Policy-Based Access Control · Policy-as-code

PBAC (Policy-Based Access Control) moves the authorization decision into an external policy engine expressed as code (such as OPA/Rego or AWS Cedar), and in practice can express RBAC, ABAC, and ReBAC rules.

Role-Based Access Control

RBAC (Role-Based Access Control) grants permissions to roles and assigns roles to users, so access is decided by the role a user holds; it is the most widely deployed authorization model.

Relationship-Based Access Control · Zanzibar

ReBAC (Relationship-Based Access Control) decides access by traversing relationships in a graph, the model popularized by Google Zanzibar and used for sharing, hierarchy, and nested ownership.

Security Assertion Markup Language · SAML 2.0

SAML (Security Assertion Markup Language) is an XML-based standard for exchanging signed authentication and authorization data between an identity provider and a service provider, and it is the incumbent protocol for enterprise single sign-on.

System for Cross-domain Identity Management

SCIM (System for Cross-domain Identity Management) is a REST API standard for automating user provisioning and deprovisioning between an identity provider and an application, keeping the account lifecycle in sync.

Web Authentication API

WebAuthn (Web Authentication) is the W3C browser API for registering and authenticating users with public-key credentials, binding every signature to the page origin to make login phishing-resistant.

AI agents

Agentic AI describes systems where a language model plans and acts, calling tools and looping until a task is done; its autonomy multiplies both value and security risk.

Token limit

The context window is the maximum amount of text, in tokens, a language model can consider at once, spanning the prompt, history, retrieved documents, and the response.

Vector embeddings

An embedding is a numeric vector that represents the meaning of text (or an image or audio) so that similar meanings sit close together, enabling search by meaning rather than exact keywords.

LoRA · Model fine-tuning

Fine-tuning adapts a pretrained model to a specific task or style by continuing training on a smaller curated dataset, updating the model's weights; it teaches behavior, not fresh facts.

LLM guardrails

Guardrails are the input, output, and system-prompt controls placed around a language model to constrain what it can receive, say, or do; each layer is porous, so they are defense-in-depth, not a fix.

Model inference

Inference is running a trained model to produce an output; it is the recurring, per-request cost of an AI product and the runtime surface attackers hit.

LLM jailbreak

A jailbreak is a prompt crafted to make a language model ignore its safety training and produce content it was tuned to refuse; it targets model alignment rather than the application's instructions.

Model Context Protocol

MCP (Model Context Protocol) is an open standard for how a language model or agent discovers and calls external tools, data, and prompts; each MCP server it connects to is a security boundary.

AI SecOps · ML security operations

MLSecOps is the practice of building security into the whole machine-learning lifecycle (data, model, deployment, and inference) as a continuous discipline, the way DevSecOps does for software.

Direct injection · Indirect injection

Prompt injection is an attack where untrusted text hijacks the instructions a language model follows; indirect injection, where hostile instructions live inside data the model reads, is the most dangerous form.

Model quantization

Quantization shrinks a model by storing its weights at lower numeric precision (for example 4-bit or 8-bit), cutting memory and inference cost at some accuracy tradeoff.

Retrieval-Augmented Generation

RAG (Retrieval-Augmented Generation) grounds a model's answer in documents retrieved at query time rather than only its training data, reducing hallucination and enabling answers over private or fresh data.

Function calling

Tool use (function calling) is the mechanism by which a language model invokes external functions instead of only generating text; it is where a language problem becomes a real-world action.

Vector store

A vector database indexes and searches embeddings by similarity rather than exact match, using approximate nearest-neighbor algorithms; it is the retrieval layer behind semantic search and RAG.

Dynamic Application Security Testing

DAST tests a running application from the outside by sending crafted requests (black-box), catching runtime and configuration flaws that static analysis cannot see; pair it with SAST.

Endpoint Detection and Response

EDR is security software on endpoints that records activity, detects malicious behavior, and lets responders investigate and contain threats, going beyond signature-based antivirus.

Static Application Security Testing

SAST analyzes an application's source code without running it to find security flaws early (white-box); its counterpart DAST tests the running app from the outside.

Security Information and Event Management

A SIEM collects, normalizes, and correlates logs and events across an environment to detect and alert on security activity; it presumes a team to run it, so it is usually premature for early startups.

Security Orchestration, Automation, and Response

SOAR automates and orchestrates the response to security alerts with playbooks, sitting downstream of detection (often a SIEM) to cut response time.

Extended Detection and Response

XDR extends EDR beyond the endpoint, correlating telemetry across endpoints, identity, network, email, and cloud into one detection and response platform.

Zero Trust Architecture · ZTA

Zero trust is a security model that trusts no user, device, or network location implicitly and verifies every request; it replaces the perimeter with per-request authentication and least privilege.

Zero Trust Network Access

ZTNA grants a user access to a specific application rather than a whole network, after verifying identity, device, and context on each connection; it is the standard replacement for the VPN.