Glossary of IAM & AI Terminology

Identity and Access Management is an industry notorious for its acronyms. Use this glossary as a quick reference guide while reading.

Agentic AI: An advanced form of artificial intelligence that goes beyond simple chatbot interactions. Agentic AI can make autonomous decisions, chain API calls, and execute complex workflows on behalf of a user to achieve a high-level goal.

AiTM (Adversary-in-the-Middle): A sophisticated phishing attack where a hacker places a reverse proxy server between the user and the legitimate login page to intercept live session cookies/tokens, bypassing traditional SMS or push-based Multi-Factor Authentication.

CAEP / SSF (Continuous Access Evaluation Protocol / Shared Signals Framework): An open standard allowing Identity Providers to push real-time security events (e.g., session revocation, password reset, device risk spike) directly to downstream Policy Enforcement Points and SaaS apps, eliminating stateless token revocation lag.

CIEM (Cloud Infrastructure Entitlement Management): A specialized cloud security solution that continuously discovers, analyzes, and rightsizes permissions and cross-account privilege escalation risks across multi-cloud environments (AWS, Azure, GCP).

Chunk-Level Security (CLS): Applying granular metadata Access Control Lists (ACLs) directly to individual text chunks within a vector database, ensuring fine-grained authorization within documents containing mixed classification tiers.

Confused Deputy Problem: A security vulnerability where a privileged entity (such as an AI Agent or service account) is tricked by an untrusted party (e.g., via Prompt Injection) into misusing its legitimate ambient authority to perform unauthorized actions.

Continuous Certification (Continuous Review): An event-driven approach to identity governance. Instead of overwhelming managers with massive, annual spreadsheet audits, access reviews are triggered dynamically whenever an identity's role, reporting line, or risk context changes.

CSPM (Cloud Security Posture Management): Automated tooling that assesses multi-cloud infrastructure configurations against compliance benchmarks (CIS, NIST) to detect misconfigurations and exposed buckets.

DLS (Document-Level Security): Applying Access Control Lists (ACLs) to entire files or data objects, ensuring that search engines and AI models only retrieve documents the querying user is permitted to see.

DPoP (Demonstrating Proof-of-Possession / RFC 9449): An application-level cryptographic mechanism that binds OAuth 2.0 access and refresh tokens to a client's private key (held in a hardware TPM), preventing stolen tokens from being replayed on external attacker machines.

Ephemeral Token: A dynamic, short-lived credential generated on the fly for a single transaction or narrow time window (e.g., 5–15 minutes) that automatically self-destructs, eliminating static hardcoded passwords.

FIDO2 / WebAuthn: An open authentication standard that enables passwordless, phishing-resistant logins by cryptographically binding credentials to the exact website origin using hardware security keys (e.g., YubiKey) or platform biometrics.

HITL (Human-in-the-Loop): An architectural circuit breaker ensuring that an autonomous AI agent can investigate and draft a high-risk mutation (e.g., firewall changes, financial transfers), but cannot execute it without explicit human approval.

IAM (Identity and Access Management): The overarching framework of policies and technologies ensuring that the right entities (humans, bots, or AI) have the right access to the right resources at the right time, for the right reasons.

IGA (Identity Governance and Administration): The policy and compliance engine of IAM. IGA tools manage identity lifecycles (JML), enforce Separation of Duties (SoD), and orchestrate access certification campaigns.

ISO/IEC 42001: The international certifiable standard establishing requirements for an Artificial Intelligence Management System (AIMS) within organizations.

ITDR (Identity Threat Detection and Response): A dedicated security layer focused on detecting and mitigating active attacks against identity infrastructure (e.g., Active Directory, cloud IdPs), such as Kerberoasting, Golden SAML, and credential stuffing.

JIT (Just-In-Time Access): A security protocol where administrative privileges are revoked by default. Admins check out elevated permissions dynamically for a specific task, which automatically expire after a few hours.

JML (Joiner, Mover, Leaver): The three fundamental phases of an employee's lifecycle. Automating JML processes ensures users get access quickly on Day 1 (Joiner), lose unnecessary access when changing roles (Mover), and are instantly locked out upon termination (Leaver).

MCP (Model Context Protocol): An open standard establishing strict, declarative JSON-RPC schema contracts and authorization boundaries between LLMs and external tools/databases.

NHI (Non-Human Identity): Any digital identity that does not belong to a human person, including service accounts, API keys, CI/CD pipelines, background scripts, and AI agents.

OIDC (OpenID Connect): An identity layer built on top of the OAuth 2.0 protocol allowing clients to verify an end-user's identity based on authentication performed by an Authorization Server.

Passkeys (Device-Bound vs. Synced): Cryptographic WebAuthn credentials. Synced Passkeys synchronize across personal cloud keychains; Device-Bound Passkeys are locked inside dedicated hardware TPMs/enclaves and cannot be exported.

PBAC (Policy-Based Access Control): An authorization model expressing access rules as declarative code (e.g., Open Policy Agent Rego, AWS Cedar), decoupling policy from application logic.

Pre-Filtering (RAG): Filtering vector database search spaces by metadata ACL tags before computing vector distance math, eliminating Denial of Service and inference leakage risks found in Post-Filtering.

Privilege Attenuation (Scope Attenuation): The cryptographic reduction of permissions when exchanging or delegating tokens (e.g., via RFC 8693) so that downstream agents or tools receive only the minimal micro-scope needed for a single task.

Prompt Injection: A cyberattack unique to LLMs where an attacker embeds malicious instructions within untrusted data, hijacking the agent's reasoning loop (the Confused Deputy).

RAG (Retrieval-Augmented Generation): Connecting an LLM to an external vector knowledge base to retrieve company-specific data and ground model responses.

RBAC (Role-Based Access Control): A coarse-grained authorization model where access is granted based on job functions (e.g., "Developer").

ReBAC (Relationship-Based Access Control): A fine-grained authorization model (pioneered by Google Zanzibar) that evaluates permissions based on a graph of relationships between subjects and resources.

RFC 8693 (OAuth 2.0 Token Exchange): The IETF standard defining how security tokens are securely traded between callers, identity providers, and autonomous agents while preserving delegation audit trails.

Secret Zero: The fundamental dilemma of how an automated workload securely authenticates to a Secrets Manager in the first place without using a hardcoded static credential.

SoD (Separation of Duties): A governance principle ensuring that no single individual or agent possesses enough permissions to execute a fraudulent or catastrophic action unilaterally.

SPIFFE / SPIRE: The Secure Production Identity Framework for Everyone. An open-source standard for issuing short-lived cryptographic X.509/JWT SVID certificates to workloads based on kernel attestation, enabling automated mTLS without static passwords.

Vector Database: A specialized database designed to store unstructured data as mathematical embedding vectors for semantic similarity search.

ZTNA (Zero Trust Network Access): A modern security architecture replacing perimeter VPNs with continuous, context-aware policy enforcement points granting micro-tunnels to individual applications.