Glossary
Plain-English definitions, with examples that connect the terminology to everyday security decisions.
AI agent
An AI-based system that can use tools or connected services to take actions toward a task.
In practice: A purchasing agent can read a catalog and prepare an order. Its permissions determine whether it can submit that order.
OWASP: Excessive AgencyAuthorization
The decision about whether an identity is allowed to access a resource or perform a particular action.
In practice: Being signed in does not automatically authorize an assistant to issue a refund.
NIST: AuthorizationExcessive agency
A risk created when an AI system has more functionality, permission, or autonomy than its task needs, enabling harmful actions.
In practice: A summarization assistant can also delete messages, even though deleting them is unnecessary for its task.
OWASP: Excessive AgencyIndirect prompt injection
An attempt to redirect AI behavior through instructions embedded in content the system reads, such as documents, webpages, or tool results.
In practice: An uploaded support document contains instructions telling the assistant to send private account information elsewhere.
OWASP: Prompt InjectionLeast privilege
Give a person or system only the access needed to perform its assigned task.
In practice: An agent that prepares orders can create drafts but cannot change supplier bank details.
NIST: Least PrivilegeOWASP
The Open Worldwide Application Security Project: an organization that publishes open resources for improving software security.
In practice: OWASP’s AI security guidance helps teams identify risks and consider safeguards.
OWASP FoundationPrompt injection
An attempt to redirect an AI system’s behavior through instructions supplied in its inputs.
In practice: A user or retrieved page tells an assistant to ignore its intended instructions and disclose information.
OWASP: Prompt InjectionSandbox
A restricted environment that limits what software can access or change while it runs.
In practice: Test a hostile document against a support assistant in an isolated environment with no access to real customer accounts.
NIST: SandboxThreat modeling
A structured way to identify what could go wrong in a system and decide which safeguards are needed.
In practice: Map where outside documents enter an assistant’s workflow and which connected tools could act on their contents.
Related guidance: OWASP Prompt Injection PreventionTrust boundary
A point where data or actions cross between parts of a system with different levels of trust or authority.
In practice: An external document is information to read; it does not have the authority of an approved instruction.
Related guidance: OWASP Prompt Injection PreventionThese definitions are plain-English teaching explanations. Consult the linked guidance for its fuller context.