Glossary · 36 terms
GenAI security glossary: the terms behind AI controls
In short
Definitions of the terms vendors and this site use when they talk about controlling generative AI, in plain English. Each term links to the explainer or approach page that goes deeper.
A
- A2A (agent-to-agent)
- Communication in which one AI agent sends requests to another. Zscaler names A2A brokers among its agent controls. Read more
- Agentic AI
- AI systems that plan and take actions, such as calling tools, editing files or sending messages, rather than only answering questions. Read more
- AI agent
- A program that uses a language model to decide which actions to take and then takes them through tools or APIs. Read more
- AI gateway
- A control point that AI requests are routed through, where policy, inspection, logging and routing to models are applied. Read more
- AI red teaming
- Testing an AI app or agent with attack prompts and scenarios to find weaknesses before attackers do. Read more
- Anonymization
- Replacing identifying data in a prompt with stand-in values so the rest can be processed. Cato lists Anonymize as a policy action. Read more
- Approved-tools list
- The published list of AI tools a company allows, with what each may be used for. Read more
C
- CASB
- Cloud access security broker: a control that gives visibility into and policy over the use of SaaS applications.
- Coaching
- A response that warns the person about a risky action and lets them decide whether to continue. Read more
- Coding agent
- An AI agent that works in a terminal or code editor, reading and changing code and running commands. Examples named on vendor pages include Claude Code, Cursor and Codex. Read more
D
- Data classification
- Labelling data by sensitivity, such as public, internal or restricted, so policy can treat each class differently.
- DLP
- Data loss prevention: controls that detect sensitive data in transit or at rest and stop or record its movement. Read more
F
- Fail closed
- Behaviour in which traffic is stopped when the security control is unavailable.
- Fail open
- Behaviour in which traffic passes uninspected when the security control is unavailable.
- FedRAMP
- The US government programme that authorizes cloud services for use by federal agencies. An authorization covers named services, not a whole company. Read more
- FWaaS
- Firewall as a service: firewall functions delivered from a provider's cloud instead of an appliance.
G
- GenAI app control
- Policy that allows, limits or blocks the use of specific generative AI applications, often by user group or action such as upload.
- Guardrails
- Runtime checks on prompts and responses that block or modify unsafe content, such as prompt injection or data leaks. Read more
I
- Inline enforcement
- A control placed in the path of traffic so it can stop or change a request before it arrives. Read more
J
- Jailbreak
- A prompt crafted to make a model ignore its instructions or safety rules.
M
O
- Out-of-band monitoring
- Watching copies of traffic, logs or APIs after the fact, without sitting in the path. Read more
P
- PoP
- Point of presence: a location where a cloud security provider receives and inspects traffic. Vendors count PoPs on different bases.
- Prompt
- The text, files or instructions sent to an AI model.
- Prompt DLP
- Data loss prevention applied to prompts and AI responses. Read more
- Prompt injection
- Instructions hidden in content an AI reads, such as a web page or a document, that try to change what the AI does. Read more
R
- Redaction
- Removing or masking sensitive parts of a prompt so the rest can be sent. Read more
S
- SASE
- Secure access service edge: networking (such as SD-WAN) and security (such as SWG, CASB, ZTNA and firewalling) delivered together as a cloud service. Read more
- SD-WAN
- Software-defined wide area network: software that routes traffic between sites and clouds over several kinds of links.
- Shadow AI
- Use of AI tools, AI features or agents that the security team has not approved or cannot see. Read more
- Single-pass inspection
- An architecture in which traffic is decrypted and inspected once for several security functions. Cato uses this wording for its engine.
- SSE
- Security service edge: the security half of SASE, typically SWG, CASB, ZTNA and DLP delivered from a cloud. Read more
- SWG
- Secure web gateway: a control that inspects and filters users' web traffic.
T
- Tool call
- A request from an AI agent to run a specific tool with specific arguments, such as reading a file or running a command. Read more
Z
- ZTNA
- Zero trust network access: access to private applications granted per user and per app after verification, instead of network-wide access.