AI Security Glossary

Search our knowledge base to break down complex AI terms and jargon into clear, practical definitions — many paired with short explainer videos for quick learning. Use it to understand key concepts, stay current with regulations, and navigate AI security with confidence.

Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
Showing 00 of 00 terms

Term

Short Description

PointGuard

Video

Malicious MCP Server

A malicious MCP server impersonates a legitimate service or smuggles attacker-controlled tools into an agent's environment. Real-world incidents from Postmark impersonation to OpenClaw skills illustrate how a single rogue server can deliver code execution, credential theft, or data exfiltration at agent speed.

Protection

Many-Shot Jailbreaking

Many-shot jailbreaking is a jailbreak technique in which an attacker provides many examples of disallowed responses inside the model's context window, exploiting in-context learning to coerce the model into similar outputs. The technique was introduced by Anthropic researchers and is widely studied.

Protection

MCP Authentication

MCP authentication is the process of verifying the identity of MCP servers, clients, and the agents calling them. It is the foundation of trustworthy MCP deployments and the most common control gap in real-world MCP incidents.

Feature

MCP Client

An MCP client is the component inside an AI agent or application that connects to MCP servers and invokes their tools. Examples include Claude Code, Cursor, GitHub Copilot CLI, and bespoke agent runtimes. The client is the trust anchor on the agent side of every MCP interaction.

MCP Discovery

MCP Discovery is the continuous practice of identifying every Model Context Protocol server, tool, and connected agent across an organization's environments. It is the inventory foundation for governing MCP risk and the prerequisite for any registry, policy, or runtime control.

Feature

MCP Gateway

An MCP gateway is a security and policy intermediary that sits between AI agents and the MCP servers they call. It centralizes identity, authorization, observability, and threat detection for every tool invocation, turning a fragmented MCP ecosystem into a governed control plane.

Feature

MCP Registry

An MCP registry is a curated catalog of MCP servers and tools used to discover, vet, and govern integrations across the agent ecosystem. Public and enterprise-internal registries are emerging as the canonical control point for managing MCP supply chain risk.

Feature

MCP Rug Pull

An MCP rug pull is an attack in which a Model Context Protocol server or tool behaves safely when first reviewed and approved, then silently changes its tool definitions or behavior to become malicious after it has earned trust.

Protection

MCP Security

MCP security is the discipline of protecting Model Context Protocol deployments end to end, covering servers, clients, transports, and the tools agents call through MCP. It addresses authentication, authorization, transport integrity, and supply chain risk across the rapidly growing MCP ecosystem.

Feature

MCP Server

An MCP server exposes tools, resources, and prompts to AI agents through the Model Context Protocol. It is the unit of integration in the MCP ecosystem and the place where most authentication, authorization, and tool-poisoning risks materialize.

MCP Tool

An MCP tool is a callable action exposed by an MCP server that an AI agent can invoke with structured arguments. Tools are the atomic primitive of agentic AI: every meaningful agent action ultimately reduces to one or more tool calls against trusted or untrusted MCP servers.

Membership Inference Attack

An attack that attempts to determine whether specific data was part of an AI model’s training set—potentially exposing sensitive or private information about individuals or proprietary records.

Protection

Memory Poisoning

Memory poisoning is an attack that plants false or malicious information in an AI agent's long-term memory, so the corrupted content influences future sessions, decisions, and actions long after the original attack.

Protection

Mission-Based Authorization

Mission-based authorization is an approach to securing AI agents in which each agent operates under a defined mission that specifies its goal, allowed tools, data, and destinations, and every action is validated against that mission before it executes.

Feature

MIT AI Risk Repository

A research initiative from MIT that catalogs emerging risks, incidents, and best practices related to AI systems—supporting safe deployment and policy development through shared, evidence-based insights.

MITRE ATLAS

A knowledge base that catalogs the tactics, techniques, and case studies of adversarial attacks on machine learning (ML) and artificial intelligence (AI) systems to help organizations understand and mitigate these threats.

MITRE ATT&CK

Knowledge base of adversary tactics and techniques based on real-world observations, used as a foundation for developing threat models and methodologies in the cybersecurity community.

MLOps

The practice of deploying, managing, and monitoring machine learning models in production to ensure they operate efficiently and effectively.

Model Card

A model card is a structured document that describes an AI model's intended use, training data, performance, limitations, and ethical considerations. Originally proposed by Mitchell et al., model cards have become a de facto standard for model documentation referenced in audits and regulatory submissions.

Model Context Protocol (MCP)

A structured format for managing inputs, context, and control signals passed to AI models—enabling more secure, transparent, and auditable language model interactions.

Model Deployment

The process of integrating a machine learning model into a production environment where it can make predictions on new data and deliver business value.

Model Discovery

Model discovery is the AI-discovery sub-discipline focused specifically on identifying every machine learning model in use across an organization, including hosted, self-hosted, fine-tuned, and embedded models. It is the prerequisite for any model-centric governance or risk program.

Feature

Model Drift

A condition where an AI model’s performance degrades over time due to changes in data, user behavior, or environment—resulting in reduced accuracy, reliability, or alignment with business goals.

Protection

Model Extraction

An attack where adversaries attempt to replicate or reverse-engineer a deployed model by querying it—stealing intellectual property or exposing sensitive model logic.

Protection

Model Fuzzing

Testing technique used to identify vulnerabilities and weaknesses in machine learning models by inputting random, unexpected, or malformed data to observe how the model responds.

Feature

Model Inference

The process of using a trained AI model to generate predictions, classifications, or responses based on new input data. Inference occurs during model deployment and often in real-time production environments.

Model Inversion

An attack where adversaries reconstruct sensitive training data by analyzing a model’s outputs—potentially exposing private or proprietary information used during model training.

Protection

Model Monitoring

Model monitoring tracks the performance, accuracy, and behavior of machine learning models in production. It helps detect drift, degradation, or security incidents—ensuring that deployed models remain effective, trustworthy, and compliant with business or regulatory expectations over time.

Feature

Model Poisoning

AI model poisoning is the manipulation of a trained model's weights or behavior so that it produces attacker-chosen outputs under specific conditions. It is distinct from training data poisoning in that the attack targets the model artifact itself, often via supply chain channels.

Protection

Model Scanning

Systematically examining AI models for security flaws, policy violations, or performance issues. It identifies risks such as embedded secrets, adversarial behavior, or unauthorized code—ensuring models are safe to deploy and comply with governance requirements.

Feature

Model Serving

The process of deploying machine learning models into production environments where they can process real-time data and generate predictions or insights.

Model Supply Chain Security

Protects the integrity of machine learning models and their dependencies—ensuring that models, datasets, libraries, and infrastructure have not been tampered with, poisoned, or sourced from untrusted repositories.

Feature

Model Weight Exfiltration

Model weight exfiltration is the theft of a trained AI model's parameters, which can let attackers copy proprietary capabilities, remove safety controls, or deploy the model for misuse without the developer's oversight.

Protection

Multi-Agent System Security

Multi-agent system security protects architectures in which multiple autonomous AI agents collaborate to complete a task. It adds new failure modes such as cross-agent prompt injection, cascading goal drift, and trust boundary erosion that single-agent security models do not address.

Feature

Multi-Turn Attack (AI)

Multi-turn attacks are AI manipulation techniques that unfold across multiple conversation turns rather than a single prompt. Attackers split malicious intent across exchanges, gradually steering the model past safety controls that only evaluate inputs in isolation.

Protection

National Institute of Standards and Technology (NIST)

An agency of the US Department of Commerce whose mission is to promote American innovation and industrial competitiveness

National Vulnerability Database (NVD)

U.S. government repository of standards-based vulnerability management data represented using the Security Content Automation Protocol (SCAP).

NIST AI 600-1

NIST AI 600-1 is the Generative AI Profile of the NIST AI Risk Management Framework. It tailors the AI RMF functions, categories, and outcomes to risks unique to generative AI, including hallucination, prompt injection, and supply chain exposure.

NIST AI Risk Management Framework (RMF)

A U.S. government framework developed by NIST to guide organizations in managing risks associated with AI systems—emphasizing trustworthiness, accountability, and governance across the AI lifecycle.

Non-Human Identity (NHI)

Non-human identity (NHI) refers to the identifiers and credentials used by software actors such as services, workloads, bots, and AI agents. As agents proliferate, NHI has become the fastest-growing identity category and now has its own dedicated OWASP Top 10.

On-Behalf-Of (OBO) Tokens for Agents

On-Behalf-Of (OBO) tokens are OAuth-style credentials that let an AI agent act on a user's behalf, carrying the user's identity through to downstream APIs while constraining scope, duration, and audience. They are the foundation of accountable agent action in enterprise environments.

Feature

Open Source Software (OSS)

Software that is released with a license allowing anyone to view, modify, and distribute the source code.

OWASP Agentic Skills Top 10

The OWASP Agentic Skills Top 10 is an OWASP project that identifies the most critical security risks in AI agent skills, the packaged instructions, scripts, and resources that extend what agents can do.

OWASP AIVSS

The OWASP AI Vulnerability Scoring System (AIVSS) is a framework for scoring the severity of security risks in AI and agentic systems, extending traditional vulnerability scoring with factors such as autonomy, tool use, and multi-agent interactions.

OWASP ASI01: Agent Goal Hijack

Agent goal hijack is an attack in which adversarial input redirects an autonomous AI agent away from its authorized objective toward an attacker-chosen goal. It is the agentic evolution of prompt injection and is catalogued as ASI01 in the OWASP Agentic Top 10.

Protection

OWASP ASI02: Tool Misuse

Tool misuse is the agent-era attack class in which an AI agent is induced to call legitimate tools in unauthorized or destructive ways. Catalogued as ASI02 in the OWASP Agentic Top 10, tool misuse is the most-cited agent failure pattern in 2026 incident reports.

Protection

OWASP ASI03: Identity and Privilege Abuse

OWASP ASI03 Identity and Privilege Abuse exploits dynamic trust and delegation in AI agents to escalate access and bypass controls. It manipulates delegation chains, role inheritance, and cached credentials so an agent's authority is reused beyond its intended scope.

Protection

OWASP ASI04: Agentic Supply Chain Vulnerabilities

OWASP ASI04 Agentic Supply Chain Vulnerabilities arise when models, tools, MCP servers, plugins, datasets, or peer agents loaded at runtime have been tampered with or compromised. Unlike static software supply chains, agentic systems compose capabilities live, expanding the attack surface.

Protection

OWASP ASI05: Unexpected Code Execution

OWASP ASI05 Unexpected Code Execution describes how agentic systems generate or invoke code that escalates into remote code execution, sandbox escape, or host compromise. Prompt injection, tool misuse, or unsafe serialization can convert text into unintended executable behavior at runtime.

Protection

OWASP ASI06: Memory and Context Poisonin

OWASP ASI06 Memory and Context Poisoning corrupts stored or retrievable agent context (memory, summaries, embeddings, RAG stores) with malicious data, so future reasoning, planning, or tool use becomes biased, unsafe, or aids exfiltration across sessions.

Protection

OWASP ASI07: Insecure Inter-Agent Communication

OWASP ASI07 Insecure Inter-Agent Communication arises when messages between AI agents lack authentication, integrity, or semantic validation, letting attackers intercept, spoof, replay, or manipulate exchanges across transport, routing, discovery, and semantic layers.

Protection

OWASP ASI08: Cascading Failures

OWASP ASI08 Cascading Failures occur when a single fault (hallucination, malicious input, poisoned memory, or corrupted tool) propagates across autonomous agents at machine speed, compounding into systemwide harm that human checks cannot catch in real time.

Protection

OWASP ASI09: Human-Agent Trust Exploitation

OWASP ASI09 Human-Agent Trust Exploitation occurs when adversaries or misaligned designs exploit human trust in AI agents (built on fluency, authority bias, and anthropomorphism) to influence approvals, extract information, or steer high-impact actions through manipulated explanations.

Protection

OWASP ASI10: Rogue Agents

OWASP ASI10 Rogue Agents are malicious or compromised AI agents that deviate from their intended function or authorized scope. They act harmfully, deceptively, or parasitically within multi-agent or human-agent ecosystems, often through emergent behavior that no single action reveals.

Protection

OWASP Top 10

A standards list from the non-profit Open Worldwide Application Security Project representing a broad consensus about the most critical security risks to web applications.

OWASP Top 10 for Agentic Applications

The OWASP Top 10 for Agentic Applications is a community-vetted standard published by the OWASP GenAI Security Project that catalogs the ten most critical security risks for autonomous AI agents, numbered ASI01 through ASI10. It is the canonical reference for agentic AI threat modeling.

OWASP Top 10 for LLM Applications (2025)

List of the most critical security risks associated with large language models (LLMs), providing guidance for identifying, understanding, and mitigating these vulnerabilities to enhance the security and integrity of AI systems.

Video

Package Manager

A tool that automates the process of installing, upgrading, configuring, and removing software packages.

Patch Management

The process of distributing and applying updates to software to fix vulnerabilities and improve functionality.

Penetration Testing

A simulated cyberattack against an application to identify security weaknesses that could be exploited by malicious actors.

Personally Identifiable Information (PII)

Information that can be used on its own or with other information to identify, contact, or locate a single person, or to identify an individual in context.

Plugin Supply Chain (AI Agents)

The AI agent plugin supply chain is the ecosystem of plugins, skills, extensions, and marketplaces that extend AI agents. Weaknesses anywhere in it can let attackers deliver malicious code that runs with an agent's privileges.

Protection

Prompt Injection

A manipulation technique where attackers insert crafted instructions into prompts to override, redirect, or subvert the intended behavior of large language models and other AI systems.

Protection

Prompt Leakage

An issue where parts of a system prompt or hidden instructions are unintentionally revealed in model outputs—exposing internal logic, rules, or security configurations.

Protection

Prompt Obfuscation

A technique used to disguise or encode malicious prompts to evade content filters, safety systems, or moderation tools in language models and AI-driven applications.

Protection

Promptware

Promptware is malicious content built from natural-language instructions rather than code, designed to hijack AI systems and carry out multi-stage attacks such as persistence, privilege escalation, and data exfiltration through the AI's own capabilities.

Protection

Provenance

The history and origin of software components, tracking their creation, modification, and distribution.

Regulatory Compliance

The act of following relevant laws, regulations, and guidelines set by governing bodies to operate within legal frameworks.

Regulatory Framework

A set of standards and principles that guide the operations and governance of an organization within its industry.

Ready to get started?

Our expert team can assess your needs, show you a live demo, and recommend a solution that will save you time and money.