AI Security Glossary

A tactic where attackers craft inputs designed to fool AI models into making incorrect predictions. These subtle manipulations pose serious risks to model reliability, accuracy, and trust.

Agent authentication is the process of verifying the identity of an autonomous AI agent before it is allowed to act, call tools, or access data. It pairs with agent identity and agent authorization to form the identity triad for agentic systems.

Agent authorization is the process of deciding, at every action, whether an autonomous AI agent is permitted to invoke a given tool, access a given resource, or perform a given operation. It applies policy at the level of intent and tool call, not just the network or API surface.

An agent covert channel is an unapproved communication path that AI agents, or attackers using them, create through shared systems such as caches, repositories, or metadata, allowing data or instructions to move between sessions, agents, or accounts that should be isolated.

Agent credential theft is the stealing of tokens, API keys, and configuration files that AI agents use to access systems, often by infostealer malware, giving attackers the agent's access without needing to bypass multi-factor authentication.

Agent discovery is the practice of identifying every autonomous AI agent operating across an organization, including developer agents, customer-facing agents, internal automations, and third-party agents embedded in SaaS apps. It is the agentic counterpart to traditional asset discovery.

Agent governance is the framework of policies, controls, and oversight that defines how autonomous AI agents are deployed, authorized, monitored, and retired. It extends traditional IT and AI governance to cover identity, authorization, behavior, and accountability for every agent in production.

Agent hijacking is the adversarial takeover of an autonomous AI agent, in which an attacker gains control of the agent's reasoning, tools, or credentials. NIST treats agent hijacking as a distinct adversarial machine learning threat that covers a broader scope than goal-level redirection.

Agentic AI Security is the discipline of protecting autonomous AI agents that perceive, plan, and act on enterprise systems. It combines identity, authorization, and runtime controls so agents can use tools, access data, and call APIs without exceeding their intended scope or trust boundary.

The protection of autonomous AI systems—such as agents or AI-driven workflows—that independently perform tasks, access tools, or make decisions on behalf of users.

An agentic browser is a web browser with a built-in AI agent that can read pages, navigate, fill forms, and take actions for the user. Examples include Perplexity Comet and AI features in Chrome and Edge. Agentic browsers create new risks because the agent acts within the user's signed-in sessions.

Agentic endpoint security is the discovery, monitoring, and control of AI agents running on employee devices, including coding agents, agentic browsers, desktop assistants, and local MCP servers.

Agent identity is the verifiable identifier and credential set assigned to an autonomous AI agent so its actions can be authenticated, authorized, and audited. Like user or service identities, agent identity binds an actor to permissions, sessions, and accountability across systems.

Agent observability is the ability to trace and understand what AI agents are doing, including their prompts, reasoning steps, tool calls, data access, and outputs, across entire workflows. It is the foundation for detecting misuse, debugging failures, and proving compliance.

An agent registry is an authoritative inventory of the AI agents operating in an organization, recording each agent's owner, purpose, identity, permissions, tools, and status. It is the system of record for agent governance.

An agent sandbox is an isolated execution environment that confines what an autonomous AI agent can do, see, and reach. Sandboxing limits blast radius when prompt injection or misalignment causes the agent to attempt destructive or out-of-scope actions.

Agent sprawl is the uncontrolled growth of AI agents across an organization's teams, platforms, and endpoints, creating overlapping, unowned, and poorly secured agents that security teams cannot see or manage.

An agent swarm is a large group of AI agents working in parallel and coordinating toward shared goals. Swarms can multiply productivity, but when they coordinate outside approved channels they can also amplify attacks, as in the 2026 OpenAI agent breach of Hugging Face.

Agent-to-agent communication describes the protocols and patterns by which autonomous AI agents exchange messages, tasks, and tool results with each other. As multi-agent architectures proliferate, A2A communication becomes both a productivity primitive and a fast-growing attack surface.

Autonomous or semi-autonomous systems that use AI to perceive, reason, and act toward goals—often coordinating across tools, APIs, or environments to perform complex, multi-step tasks on behalf of users.