Glossary · 127 terms
AI security glossary and the identities behind it
Plain answers to the questions people ask about AI security: AI apps and agents, MCP, prompt injection and other AI attacks, AI data loss, and the standards that frame them. Then the human, non-human and AI-agent identities that every one of those depends on.
Weighing two categories? See the AI security comparisons.
Start here
Matches term names and definitions. Press slash to focus search, Escape to clear, Enter to jump to the first result./
127 terms
AI security fundamentals
- What is AI security?AI security covers chatbots, coding agents and connected AI apps at work. See where data leaks, how agents get misused and which controls actually hold.
- What is generative AI security?Generative AI security covers ChatGPT, Microsoft 365 Copilot, coding assistants and custom apps. The real risks, how controls differ by use, and what works.
- What is LLM security?LLM security protects language models, the apps built on them and their data. The main threats, how an attack on a support assistant works, and what stops it.
- What is AI runtime security?AI runtime security checks live prompts, uploads, tool calls and agent actions against policy. Learn where controls belong and how to prove that they work.
- What is AI for security vs security for AI?AI for security uses AI to do security work. Security for AI protects the AI your company uses. Compare goals, examples, testing and how to evaluate tools.
- What are AI security tools?AI security tools discover, test and control risk in AI apps, agents and employee AI use. Compare the main types, where each sits, and what to test first.
- What are AI security best practices?AI security best practices for enterprises: find the AI already in use, keep sensitive data out, limit agent permissions, test controls, rehearse response.
- What is zero trust for AI?Zero trust for AI checks every agent action, well past the login. How to scope permissions, check tool calls, limit delegation and prove the controls work.
- What is AI data privacy?AI data privacy covers what AI tools collect, keep and share. Why no-training is not no-storage, why the signed-in account matters and how to limit uploads.
- What is responsible AI?Responsible AI turns fairness, privacy, safety and accountability into checks you can run. Who owns it, how it differs from governance and how to test it.
AI security programs and controls
- What is an AI security program?An AI security program assigns owners, policy, access limits and incident response for enterprise AI. How to build the inventory, run a pilot and test it.
- What is AI usage control?AI usage control sets rules for employee and agent AI use. Learn how to govern accounts, protect sensitive data, limit actions, and test enforcement.
- What is AI security posture management (AI-SPM)?AI-SPM finds the AI apps, agents and grants in your environment, checks what each can reach and ranks the fixes. Here is what to inventory and check.
- What is AI detection and response (AIDR)?AI detection and response finds risky AI use and agent actions, works out what really happened and contains it. How to scope incidents and pick a response.
- What is AI data loss prevention (AI DLP)?AI DLP stops sensitive data leaving through prompts, uploads, pastes and AI connectors. Learn when to mask, block or warn, and where the gaps usually are.
- What is DSPM for AI?DSPM for AI finds sensitive data that Copilot, agents and AI apps can reach but shouldn't. Learn how to trace the access path, fix oversharing and prove it.
- What is an AI gateway?An AI gateway sits between your apps and model providers to route requests and apply policy. Learn how it works, what it misses and how to deploy it safely.
- What is an AI firewall?An AI firewall inspects prompts and responses and blocks or redacts what policy forbids. How it works, where coverage stops and how to test it first.
- What is an AI audit trail?An AI audit trail links AI requests, identities, approvals and actions. What to record, how to protect it and how to use it when an agent does something odd.
- What is AI agent observability?AI agent observability lets you rebuild what an agent did: its prompts, tool calls, identity and results. See how security tracing differs from dev tracing.
- What is AI incident response?AI incident response covers data leaks and hijacked agents. What to check first, which evidence to keep, how to revoke access and when to restore safely.
- What is AI red teaming?AI red teaming is authorized attack testing of AI systems. See how to scope a test, run a prompt injection scenario, record evidence and fix what breaks.
- What is an AI bill of materials (AI-BOM)?An AI bill of materials records the models, datasets, packages and MCP servers in an AI system. What to include, how it differs from an SBOM, how to use it.
- What is AI risk assessment?AI risk assessment weighs the data, access, autonomy and possible harm of one AI use case. What to check, how to test controls, and when to reassess it.
- What is AI vendor risk management?AI vendor risk management reviews how AI providers handle your data, what connectors can reach and how terms change. What to check before and after rollout.
- What is an AI acceptable use policy?An AI acceptable use policy says which AI tools, accounts and data people may use at work. What to put in it, how to enforce it and how to handle exceptions.
- What is AI security awareness training?AI security awareness training teaches staff which AI accounts to use, what data stays out of prompts and uploads, and when to stop an agent. See exercises.
- What is AI compliance?AI compliance means meeting the rules that apply to your AI use and proving it. How to scope use cases, map requirements to controls and keep audit evidence.
- What is AI governance?AI governance defines who owns AI risk, which uses are approved, and how controls are enforced. Learn practical steps for enterprise oversight and security.
AI threats and attacks
- What is prompt injection?Prompt injection redirects AI through malicious instructions. Learn how direct and indirect attacks work, why agents are exposed, and how to limit harm.
- What is indirect prompt injection?Indirect prompt injection hides instructions in emails, web pages and files an AI reads. See how it hijacks agents and how to limit what they can do.
- What is AI jailbreaking?AI jailbreaking tricks a model past its safety rules. Learn how it differs from prompt injection and how tight permissions limit damage when refusals fail.
- What is MCP tool poisoning?MCP tool poisoning hides instructions in the tool descriptions an AI agent reads. See how it turns a coding agent against you and how to limit the damage.
- What is AI agent hijacking?AI agent hijacking turns an agent's own access toward an attacker's goal. See how it happens, how it differs from prompt injection and how to contain it.
- What is agent memory poisoning?Agent memory poisoning plants false facts or instructions in what an AI agent remembers so they steer later tasks. See how it works and how to clean it up.
- What is system prompt leakage?System prompt leakage is when an AI app reveals its hidden instructions. See what actually leaks, why secrets never belong in a prompt and how to test for it.
- What is insecure AI output handling?Insecure AI output handling is when an app hands model output to a shell, browser, database or tool without checks. Examples, controls and how to test it.
- What is data and model poisoning?Data and model poisoning tampers with training data or weights to plant attacker behavior. How it works, how it differs from RAG poisoning and how to respond.
- What is model extraction?Model extraction copies a model's behavior through API queries. How it differs from stealing weights, what an attack looks like and how to protect models.
- What is AI supply chain security?AI supply chain security covers the models, packages, MCP servers, skills and AI services you depend on. Learn how attacks work and how to limit the damage.
- What is slopsquatting?Slopsquatting is when attackers register package names that AI coding tools invent. See how the attack works and how to keep developers and agents from it.
- What is RAG security?RAG security keeps retrieval-augmented AI from leaking documents or following poisoned sources. See how to enforce permissions, test retrieval and fix leaks.
- What is AI copilot oversharing?Copilot oversharing is when an AI assistant surfaces files people could open but should not see. Learn why it happens and how to fix and test the access.
- What is AI data leakage?AI data leakage exposes sensitive information through prompts, uploads, connectors, and outputs. Learn how to prevent disclosure across AI workflows.
- What is AI-powered phishing?AI-powered phishing uses generated text, voice clones and fake video to win access. See what the lures look like and which defenses work whoever wrote them.
- What is AI hallucination risk?AI hallucination risk is harm from AI output that sounds right but is invented. See where it bites in code and incident work, and how to stop it acting.
- What is deepfake fraud?Deepfake fraud uses AI-made voices or video to pose as someone you trust. How the scams work, why spotting fakes fails and which verification steps hold up.
- What is LLMjacking?LLMjacking is running AI models on someone else's stolen cloud or API credentials. How keys leak, what the abuse looks like in logs and how to contain it.
- What is AI API key security?AI API key security protects the credentials behind AI services. How keys leak through prompts and agents, how to scope and rotate them and what to revoke.
- What is shadow MCP?Shadow MCP is an MCP server no one approved, wired into Cursor, Claude or another AI tool. Learn how to find these servers, scope them and shut them off.
- What is an AI insider threat?An AI insider threat is trusted access misused through AI, by accident or on purpose. See how personal logins, connected apps and agents expose company data.
- What is adversarial machine learning?Adversarial machine learning covers attacks that fool, poison, copy or leak from AI models. Learn the attack types, what access attackers need, and defenses.
- What is a model inversion attack?A model inversion attack works backward from a model's outputs to sensitive training data. See how it differs from membership inference and how to cut risk.
- What is excessive agency?Excessive agency gives AI agents more tools, permissions, or autonomy than their tasks require. Learn the risks, enterprise examples, and practical defenses.
- What is the confused deputy problem?The confused deputy problem lets attackers misuse a trusted program's authority. Learn how it affects AI agents and how to enforce safer delegation.
AI apps, agents and usage
- What is agentic AI?Agentic AI picks its own next step toward a goal and uses tools to act on software and data. How it differs from chatbots, and what it means for security.
- What is an AI agent?An AI agent uses a model to pick and run actions toward a goal. How agents like Claude Code work, what access they need and how to stop the unsafe actions.
- What is Model Context Protocol (MCP)?Model Context Protocol links AI apps like Claude Code and Cursor to tools and data through MCP servers. How it works, where it fails and how to run it safely.
- What is account-aware AI access?Account-aware AI access applies policy to the signed-in AI account. Learn how to separate approved workspaces from personal use and handle unknown accounts.
- What is browser extension risk?Browser extension risk exposes enterprise data through browser add-ons. Learn how to review permissions, control installations, and respond to abuse.
- What is an AI agent identity?AI agent identity connects an agent to its owner, permissions, and actions. Learn how to scope access, preserve delegation, and revoke credentials safely.
- What is agentic AI security?Agentic AI security protects AI agents that use tools and take action. Learn how to limit permissions, prevent data exposure, and contain unsafe behavior.
- What is MCP security?MCP security protects AI connections to tools and data. Learn how to manage server trust, credentials, prompt injection, and sensitive tool actions.
- How does MCP authentication work?MCP authentication controls access to AI tools and data. Learn how OAuth, token validation, and least privilege help protect enterprise MCP servers.
- What is shadow AI?Shadow AI is workplace AI use outside approved oversight. Learn how to discover tools, assess data access, and guide employees toward safer AI use.
- What is AI agent delegation?AI agent delegation lets agents act on a user's or system's behalf. Learn how to limit permissions, control sub-agents, and audit delegated actions.
- What is the OAuth on-behalf-of flow for AI agents?The OAuth on-behalf-of flow lets services call APIs for a user. Learn how delegation works, where agent risks arise, and how to secure downstream access.
- What is AI agent sprawl?AI agent sprawl leaves agents, owners, and permissions untracked. Learn how to find unmanaged agents, limit access, and retire them safely across teams.
- What is an AI agent registry?An AI agent registry catalogs agents, owners, and access. Learn what to record, how to keep entries current, and where runtime security controls fit.
- What is the A2A protocol?The A2A protocol lets AI agents exchange work across platforms. Learn how discovery, authentication, delegation, and access controls shape its security.
- What are AI agent guardrails?AI agent guardrails limit what agents can access, share, and change. Learn how permissions, tool controls, and approvals reduce enterprise security risk.
- What is Microsoft Entra Agent ID?Microsoft Entra Agent ID gives AI agents managed identities. Learn how agent access, ownership, permissions, and lifecycle controls support security.
- What is AI agent authorization?AI agent authorization sets what an agent may do and under whose authority. How to scope permissions, bind human approvals and revoke access that lingers.
- What is an AI agent kill switch?An AI agent kill switch stops an agent and cuts its access. Learn why closing the chat isn't enough, what token revocation misses and how to test a shutdown.
- What are computer use agents?Computer use agents click and type through browsers and desktops like a person. Learn why signed-in sessions are the risk and how to limit what agents can do.
- What is coding agent security?Coding agent security limits what tools like Claude Code, Codex and Cursor can read, run and send. Learn how to protect secrets, repos and release workflows.
- What is vibe coding security?Vibe coding security covers the risk of shipping AI-written code with little review: leaked secrets, invented packages, weak access checks and agent access.
- What is agentic browser security?Agentic browser security controls what AI agents can read, send and change through a signed-in browser. Learn the risks, the controls and how to test them.
- What is multi-agent system security?Multi-agent system security controls what passes between AI agents, from delegated permissions to forwarded content and shared memory. Learn to contain it.
- What is AI agent sandboxing?AI agent sandboxing limits what an agent's commands can touch on disk and the network. Learn container vs VM tradeoffs and why signed-in accounts need more.
- What is AI note-taker risk?AI note-takers can expose meeting content, overshare transcripts and keep calendar access long after a call. How to control capture, sharing and permissions.
- What is AI agent skills and plugins security?Skills and plugins add instructions, code and connections to an AI agent. What can go wrong, how to review additions and how to limit what they can reach.
- What is human-in-the-loop for AI agents?Human-in-the-loop for AI agents means a person approves consequential actions before they run. Where to put the gates, who decides and how to avoid fatigue.
- What is ChatGPT security for business?ChatGPT security for business: personal vs company workspaces, what no-training really covers, custom GPT and connector risk, and how to respond to leaks.
- What is Microsoft 365 Copilot security?Microsoft 365 Copilot security starts with SharePoint permissions, sensitivity labels, agent connections and audit records. What to check before rollout.
AI standards and frameworks
- What is the OWASP LLM Top 10?The OWASP LLM Top 10 lists the main security risks in apps built on large language models. See the 2025 categories and how to test them in your own stack.
- What is the OWASP Top 10 for Agentic Applications?The OWASP Top 10 for Agentic Applications names the main ways AI agents fail. See the ten risks, how they differ from the LLM list and how to test for them.
- What is MITRE ATLAS?MITRE ATLAS catalogs how attackers go after AI systems. See how it sits next to ATT&CK and how to turn its techniques into threat models and red team tests.
- What is the NIST AI Risk Management Framework?The NIST AI RMF organizes AI risk work into Govern, Map, Measure and Manage. How each function works, what the GenAI Profile adds and where to start.
- What is ISO/IEC 42001?ISO/IEC 42001 sets requirements for an AI management system. What it covers, how it differs from 27001, what certification proves and how to prepare.
- What is the EU AI Act?The EU AI Act explained for enterprises: provider vs deployer roles, risk tiers, high-risk duties, AI literacy, and how to build an inventory that holds up.
- What is AI risk management?AI risk management ties each AI use to an owner, tested controls and a recorded call on leftover risk. How it works, what to assess and which frameworks fit.
- What is AI TRiSM?AI TRiSM is Gartner's AI trust, risk and security management framework. How it differs from governance and how to turn an AI policy into tested controls.
- What is MLSecOps?MLSecOps builds security into how ML and LLM systems are built, released and run. Learn what to check before deploying a model and what to watch afterward.
Foundations
- What is identity security?Identity security protects human, non-human, and AI-agent identities through posture, threat detection, and runtime enforcement. Definitions and key stats.
- What does zero trust mean for identity?Zero trust means every identity, human or machine, is verified continuously at each request. Learn where NHIs and AI agents break zero-trust assumptions.
- What is the principle of least privilege?Least privilege means every identity gets only the permissions its task requires. Why it decays, and how AI agents make enforcement dynamic.
- What is the difference between PAM and IAM?IAM manages access for every identity. PAM adds stricter controls for privileged accounts. Compare them side by side and see where NHIs fit.
- What is identity sprawl?Identity sprawl: the unchecked growth of human and machine identities across SaaS and cloud. What drives it, how to measure it, and how to rein it in.
- What is an identity fabric?An identity fabric weaves fragmented IAM tools into one consistent layer for human, machine, and AI-agent identities. Definition, principles, and stats.
- What is CIEM?CIEM discovers and right-sizes permissions across cloud infrastructure. How it works, how it differs from ISPM, PAM, and NHI governance, plus key stats.
- What is identity security posture management (ISPM)?ISPM continuously finds misconfigurations, dormant accounts, and excessive privilege across human and non-human identities. See how it differs from ITDR.
- What is identity threat detection and response (ITDR)?ITDR is the Gartner-coined discipline of detecting and responding to identity attacks. See how it differs from IAM and where NHIs and AI agents strain it.
Human identity threats
- What is account takeover (ATO)?What account takeover is, how attackers get in through phishing, credential stuffing, session theft, and MFA fatigue, and why detection must happen live.
- What is session hijacking?Session hijacking steals the token or cookie minted after login, bypassing MFA entirely. How AitM kits and infostealers take over authenticated sessions.
- What is token theft?Token theft steals session cookies, OAuth tokens, and API keys to bypass MFA. How the attacks work, why MFA does not stop them, and how to detect replay.
- What is consent phishing?Consent phishing explained: how malicious OAuth apps use real consent screens to grab token access that bypasses MFA, and how to detect and prevent it.
- What is phishing-resistant MFA?Phishing-resistant MFA binds logins to the real site via FIDO2/WebAuthn, defeating AitM kits that bypass OTP and push. What it stops and what remains.
Non-human and workload identity
- What is a non-human identity (NHI)?A non-human identity (NHI) is a software actor, not a credential: service accounts, workloads, and AI agents. Definitions, counts, and how to secure them.
- What is the difference between a non-human identity and a machine identity?Machine identity covers infrastructure: devices, workloads, certificates. NHI is the broader umbrella adding SaaS actors and AI agents. Taxonomy and table.
- What is a machine identity?A machine identity authenticates devices, workloads, and services using certificates and keys. Learn how it fits under NHI and why cert lifecycles hurt.
- What is a service account and why is it risky?What a service account is, why long-lived, overprivileged, unowned accounts are risky, and how discovery, ownership binding, and deprovisioning help.
- What is workload identity?Workload identity lets pods, VMs, and functions prove what they are with short-lived attested credentials instead of static secrets. SPIFFE explained.
- What is workload identity federation?Workload identity federation lets workloads authenticate across clouds by exchanging platform-issued tokens for short-lived credentials, no static keys.
- What is SPIFFE?SPIFFE is the open standard for workload identity: SPIFFE IDs, SVIDs, SPIRE attestation, and where the standard stops short for AI agents.
- What are ephemeral credentials?Ephemeral credentials are short-lived secrets issued on demand and expired automatically. How they work, common forms, and what they solve vs static keys.
- What is API key security?API key security explained: why static bearer keys leak from repos, logs, and AI prompts, and how scoping, rotation, and runtime identity reduce the risk.
- What is secrets management?Secrets management stores, rotates, and injects credentials for software. Learn how vaults work, why vaulting alone fails, and how secrets tie to identity.
- What is secret sprawl?Secret sprawl is the uncontrolled spread of credentials across code, chat, CI, AI prompts, and browsers. Why every leaked secret is an identity problem.
- What is a token vault?A token vault stores and exchanges OAuth tokens so AI agents can act for users without holding credentials. How token vaults work and how to secure them.
- What is non-human identity lifecycle management?Non-human identity lifecycle management governs service accounts, keys, and AI agents from creation to decommissioning. See why offboarding is OWASP NHI1.
- What is the OWASP Non-Human Identity Top 10?The OWASP NHI Top 10 (2025) ranks the ten most critical non-human identity risks, from improper offboarding to secret leakage. Full list, one line each.
- What are OAuth app risks?OAuth app risks explained: consent phishing, scope creep, dormant grants, and invisible SaaS-to-SaaS access, plus how to find and revoke risky app tokens.
Enforcement and runtime
- What is runtime identity security?Runtime identity security protects identities while access is exercised, not just granted. How it differs from ISPM, ITDR, PAM, and posture tools.
- What is enterprise browser security?What enterprise browser security is, how extensions, enterprise browsers, and gateways compare, and why in-session visibility needs identity tie-back.
- What is just-in-time (JIT) access?Just-in-time (JIT) access replaces standing privileges with short-lived grants. How JIT works for humans, NHIs, and AI agents, and what it leaves out.
- What is zero standing privileges (ZSP)?Zero standing privileges (ZSP) removes always-on access: rights are granted just in time and revoked at completion. How it works, limits, and AI agents.
All terms, A to Z
A
- A2A protocol
- Account takeover (ATO)
- Account-aware AI access
- Adversarial machine learning
- Agent memory poisoning
- Agentic AI
- Agentic AI security
- Agentic browser security
- AI acceptable use policy
- AI agent
- AI agent authorization
- AI agent delegation
- AI agent guardrails
- AI agent hijacking
- AI agent identity
- AI agent kill switch
- AI agent observability
- AI agent registry
- AI agent sandboxing
- AI agent skills and plugins security
- AI agent sprawl
- AI API key security
- AI audit trail
- AI bill of materials (AI-BOM)
- AI compliance
- AI copilot oversharing
- AI data leakage
- AI data loss prevention (AI DLP)
- AI data privacy
- AI detection and response (AIDR)
- AI firewall
- AI for security vs security for AI
- AI gateway
- AI governance
- AI hallucination risk
- AI incident response
- AI Insider Threat
- AI jailbreaking
- AI note-taker risk
- AI red teaming
- AI risk assessment
- AI risk management
- AI runtime security
- AI security
- AI security awareness training
- AI security best practices
- AI security posture management (AI-SPM)
- AI security program
- AI security tools
- AI supply chain security
- AI TRiSM
- AI usage control
- AI vendor risk management
- AI-powered phishing
- API key security
C
I
M
N
O
S
Go deeper: AI security, built on identity, and The Non-Human Majority, Identra’s point of view on securing the identities that now outnumber us.