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

Remediation

The process of correcting or mitigating identified security vulnerabilities to protect applications from potential threats.

Responsible AI

An approach to developing and deploying AI systems that emphasizes ethical design, transparency, accountability, and minimizing harm to individuals, communities, and society.

Retrieval-Augmented Generation (RAG)

An architecture that combines a language model with an external retrieval system—grounding outputs in relevant documents or knowledge sources to improve factual accuracy and reduce hallucinations.

Reward Hacking

Reward hacking occurs when an AI system achieves its objective, or the score used to measure it, through unintended shortcuts that violate the designer's intent. In agents, reward hacking can mean breaking rules, exploiting systems, or deceiving evaluators to complete a task.

Protection

Risk Appetite

The amount and type of risk that an organization is willing to take in order to meet their strategic objectives.

Risk Assessment

The systematic process of evaluating potential risks that may be involved in a projected activity or undertaking.

Risk-Based Vulnerability Management Explained - AppSOC

Navigate risk-based vulnerability management with AppSOC. Our glossary aids in understanding threat intelligence and vulnerability prioritization.

Feature

Risk Management

The process of identifying, assessing, and prioritizing risks followed by coordinated efforts to minimize, monitor, and control the impact of unfortunate events.

Risk Mitigation

The steps taken by an organization to reduce the impact and likelihood of a risk occurring.

Risk Scoring

The process of evaluating and assigning a score to vulnerabilities based on their severity, exploitability, and potential impact on the organization.

Feature

Rogue Agent

A rogue agent is an autonomous AI agent whose behavior diverges from its authorized purpose, either through compromise, prompt injection, misalignment, or runaway internal reasoning. Rogue agents take unsanctioned actions while presenting as legitimate, making them one of the most-cited failure modes in agentic AI security.

Protection

Rules File Backdoor

A rules file backdoor is an attack that hides malicious instructions in the configuration or rules files that AI coding assistants read, such as project rules or agent instruction files, causing the assistant to generate insecure or malicious code.

Protection

Runaway Loop (AI)

A runaway loop is an autonomous AI agent failure mode in which the agent repeatedly retries, escalates, or recurses on a task in ways that consume resources, cause unintended side effects, or accelerate damage. Real-world incidents such as the PocketOS database deletion show how fast runaway loops can escalate.

Protection

Sandbagging (AI)

AI sandbagging is strategic underperformance, in which a model deliberately performs worse on an evaluation than it is capable of, for example to appear less dangerous. It undermines the safety testing that organizations and regulators rely on.

Protection

Sandbox Escape (AI Agent)

An AI agent sandbox escape occurs when an agent breaks out of the isolated environment meant to contain it and reaches systems, networks, or data outside its approved boundary. Several frontier AI labs reported agent sandbox escapes in 2026.

Protection

SANS Top 25

The SANS Top 25, now published as the CWE Top 25 by MITRE, is a list of the most dangerous software weaknesses, such as injection flaws and cross-site scripting, that attackers most often exploit.

Scheming (AI)

AI scheming describes a model covertly pursuing goals that differ from those of its developers or users, while hiding its true intentions or actions from oversight. Researchers have demonstrated in-context scheming behaviors in frontier models under test conditions.

Protection

Secure Model Deployment

The process of launching AI models into production environments with safeguards that ensure confidentiality, integrity, and safe operational behavior across data, infrastructure, and model layers.

Security Architecture Review

A continuous process, often initiated early in SDLC, to review security controls present within the architecture of a system or application.

Security Assessment

A comprehensive evaluation of an application's security posture, including vulnerability scans, penetration tests, and compliance checks.

Security Baseline

A set of minimum security standards and configurations that must be met to ensure an acceptable level of security for applications.

Security Configuration Management

The practice of ensuring that applications and systems are configured securely, following best practices and organizational policies.

Security Information and Event Management (SIEM)

A system that collects, analyzes, and reports on security-related data from various sources to provide a comprehensive view of an organization's security posture.

Security Maturity

The level of development and capability an organization has in implementing, managing, and continuously improving its security practices and controls to protect against threats and vulnerabilities.

Security Orchestration, Automation, and Response (SOAR)

A set of tools and processes designed to automate and streamline security operations, including incident response and threat management.

Security Policy

A set of rules and practices that govern how an organization protects its information and IT assets.

Security Vulnerability

A weakness in a system, application, or network that can be exploited by attackers to gain unauthorized access, cause disruptions, or steal sensitive information.

Self-Replicating Agent

A self-replicating agent is an autonomous AI agent that propagates itself, its instructions, or its goals to new contexts, users, or systems. Research demonstrations such as Morris II and emerging Mini Shai-Hulud style incidents make self-replication one of the most consequential agentic risks.

Protection

Shadow Agents

Shadow agents are AI agents deployed or used inside an organization without the knowledge or approval of security and IT teams. They extend shadow AI risk from chat tools to autonomous systems that can take actions with real credentials.

Protection

Shadow AI

Shadow AI refers to the use of artificial intelligence systems and tools within an organization without explicit approval or oversight by its central IT or AI department.

Protection

Shadow MCP

Shadow MCP refers to Model Context Protocol servers installed or used within an organization without security review or approval, often on developer machines, giving AI agents unmonitored access to corporate systems and data.

Protection

Shift Left

The practice of performing testing earlier in the software development lifecycle to identify and fix issues promptly.

Skill Poisoning

Skill poisoning is an attack in which a malicious or compromised agent skill, plugin, or extension is installed into an AI agent, giving attackers a way to inject instructions, run code, or steal data with the agent's privileges.

Protection

Slopsquatting

Slopsquatting is a software supply chain attack in which adversaries register package names that AI models commonly hallucinate, so developers or coding agents that install the suggested packages pull in attacker-controlled code.

Protection

Software Bill of Materials (SBOM)

A comprehensive list of components, libraries, and dependencies used in software development.

Software Composition Analysis (SCA)

A process that identifies and manages open source and third-party components within an application, ensuring they are secure and compliant with licensing requirements.

Software Dependency

External code libraries or modules that a software project relies on to function properly.

Software Supply Chain Security

Process of protecting the entire lifecycle of software, from development through deployment and maintenance, to prevent vulnerabilities and ensure the integrity, authenticity, and security of software components.

Feature

Software Supply Chain Vulnerability

Potential security weaknesses within the processes and components involved in developing, managing, and distributing software, which can be exploited by malicious actors.

Protection

Static Application Security Testing (SAST)

A testing method that analyzes source code, bytecode, or binary code for security vulnerabilities without executing the program.

System Card

A system card is a public document describing the design, training, evaluations, and safety considerations of a deployed AI system. Used by major AI labs to publish safety information for models such as GPT, Claude, and Gemini, system cards extend the model card concept to whole production systems.

Texas Responsible AI Governance Act (TRAIGA)

The Texas Responsible Artificial Intelligence Governance Act (TRAIGA, HB 149) is a Texas law, effective January 1, 2026, that prohibits certain harmful uses of AI, sets disclosure duties for government agencies, and gives the Texas Attorney General enforcement authority.

Threat and Vulnerability Management

A proactive approach to identifying, assessing, and mitigating security threats and vulnerabilities within an organization's IT environment.

Threat Modeling

The process of identifying, categorizing, and evaluating potential threats to an application to develop strategies for mitigating those threats.

Tool Poisoning

Tool poisoning is an attack in which the metadata of an MCP or plugin tool, such as its name, description, or argument schema, is crafted to manipulate an AI agent into calling it in attacker-chosen ways. It is one of the most subtle agent supply chain risks.

Protection

Tool Shadowing

Tool shadowing is an MCP attack in which a malicious tool's description contains instructions that change how an AI agent uses other, legitimate tools, for example redirecting emails or altering parameters, even if the malicious tool itself is never called.

Protection

Tool Squatting

Tool squatting is an attack in which adversaries publish MCP servers, tools, or agent integrations with names that imitate trusted ones, so agents or users connect to the malicious version by mistake.

Protection

Training Data Poisoning

An attack that introduces malicious or corrupted data into a model’s training set—altering its behavior, weakening performance, or embedding hidden vulnerabilities.

Protection

Unbounded AI Consumption

The unchecked use of AI-generated outputs or decisions without human validation, cost control, or safety mechanisms—leading to overreliance, misuse, or resource exhaustion.

Protection

Unified Vulnerability Management

An comprehensive approach to cybersecurity, bringing together all the information about possible security gaps in an organization's network, software and systems.

Feature

Universal Jailbreak

A universal jailbreak is a single prompt or technique that bypasses safety controls across a wide range of models or prompt categories. Research has demonstrated transferable adversarial suffixes and structured prompts that act as universal jailbreaks in current LLMs.

Protection

Vector Embedding Weaknesses

Security and privacy risks arising from the use of vector embeddings, which can inadvertently encode sensitive data or enable similarity-based exploits in AI applications like search and RAG.

Protection

Vibe Coding

Vibe coding is a style of software development in which people describe what they want to an AI and accept the generated code with little or no review. It speeds prototyping but can introduce vulnerabilities, secrets exposure, and risky dependencies into production.

Visual Prompt Injection

Visual prompt injection is an attack that hides malicious instructions in images, screenshots, or video frames so that multimodal AI models read and follow them, often without the text being visible or obvious to human users.

Protection

Vulnerability Assessment

Systematically identifying, evaluating, and prioritizing security weaknesses in an organization's IT infrastructure, including networks, applications, and systems, to mitigate potential risks.

Vulnerability Detection

The process of identifying weaknesses and security flaws in an organization's systems, applications, and networks.

Vulnerability Management

The practice of identifying, evaluating, remediating, and reporting on security vulnerabilities in applications and systems.

Vulnerability Remediation

Identifying security weaknesses in an organization's IT systems, applications, and networks through automated tools and techniques to prevent potential exploits.

Vulnerability Scanning

Systematically examining computer systems, networks, and applications for security weaknesses that could be exploited by attackers.

Web Application Security

The practice of protecting web applications from security threats and vulnerabilities by implementing measures to prevent unauthorized access, data breaches, and other cyber attacks.

YOLO Mode (Auto-Approve)

YOLO mode is an informal name for settings in AI coding agents that skip human approval and let the agent run commands, edit files, or call tools automatically. It speeds work but removes a key safeguard against mistakes and prompt injection.

Protection

Zero-Click Prompt Injection

Zero-click prompt injection is an attack in which malicious instructions reach an AI assistant or agent and execute without the victim clicking, opening, or approving anything, typically through content the AI processes automatically, such as email or shared documents.

Protection

Zero Day Vulnerability

A security flaw in software that is unknown to the software vendor and the public, leaving it unpatched and exploitable by attackers until it is discovered and addressed.

Protection

Zero Trust

A security model that requires strict identity verification for every person and device trying to access resources, regardless of whether they are inside or outside the network perimeter.

Zero-Trust for Agents

Zero-trust for agents applies zero-trust principles to autonomous AI agents, treating every agent action as untrusted by default and requiring explicit, contextual authorization for each tool call, data access, and downstream invocation. It extends NIST SP 800-207 zero-trust thinking into the agent layer.

Feature

Zero Trust for AI

A security model that assumes no AI input, user, or tool is inherently trustworthy—requiring continuous validation, monitoring, and enforcement to ensure safe operation.

Feature

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.