GLOSSARY

Endpoint AI terms: AI endpoint control in plain English

Agent Control Standard (ACS) #

An open standard for transparency and control of enterprise AI agents. The OWASP GenAI Security Project said in September 2026 that it had been donated to the project, to extend its guidance toward runtime enforcement.

Source: OWASP GenAI Security Project: Agent Control Standard · Read 27 Sep 2026

Agent hooks #

Points in an AI agent's work, such as just before it runs a tool, where an administrator's rule can check the action and let it run, change it or stop it. Lasso Security and Zenity describe controls that use them.

AI endpoint control #

Anything that sees or limits what AI tools do on, or from, an employee's device. This site sorts them into five types: on-device control, browser extension, network or SSE proxy, AI or MCP gateway, and policy only. See control types.

AI gateway #

A proxy that calls to AI models are routed through, so they can be filtered, logged and governed in one place. See the gateway guide.

AI security posture management (AI-SPM) #

Continuous checks for risky settings and excess permissions across AI models, agents and their connections, before anything happens at runtime.

AI usage control #

Gartner's name, as Harmonic Security reports it, for technology that discovers, assesses and controls how employees use AI tools.

Source: Harmonic Security: Gartner AI usage control research · Read 27 Sep 2026

AI-BOM #

An AI bill of materials: a list of the models, agents, tools and data sources a company's AI use depends on. Lasso Security uses the term for its discovery output.

Source: Lasso Security homepage · Read 27 Sep 2026

Allow, ask or deny #

Three possible answers to an AI agent action: let it run, pause and ask the person to confirm, or stop it. Bay uses these three words for its decisions.

Source: Bay homepage · Read 27 Sep 2026

Approved-tools list #

A written list of the AI tools a company allows, with who may use each and for what data. It is the core of a policy-only approach.

Browser extension (for AI control) #

Software added to the browser that watches or controls what employees type, paste and upload into AI chat tools. See the browser extension guide.

Cloud development environment #

A development workspace that runs on a remote server and is used through a browser or thin client. AI coding agents running there act on the server, not on the laptop.

Coding agent #

An AI agent built for software work, such as Claude Code, Codex or Cursor, that can read and edit code, run shell commands, install packages and connect to MCP servers.

Data loss prevention (DLP) #

Tools that detect sensitive data and stop it from leaving through email, uploads, devices or web traffic. Zscaler, for example, describes DLP blocking across GenAI apps. Many DLP products predate AI tools, so ask what they see of AI chat and agents.

Source: Zscaler: Secure the use of generative AI · Read 11 Oct 2026

Desktop AI agent #

An AI application installed on the computer, rather than used in a browser tab, that can act on files and other applications with the user's permissions.

EDR (endpoint detection and response) #

Security software on devices that detects and responds to malware and intrusion by watching processes, files and network activity. AI endpoint controls work alongside it, and some are delivered through it.

Excessive agency #

An AI agent having more tools, permissions or autonomy than its task needs, so a mistaken or manipulated agent can do real harm. It ranks third in the OWASP 2026 Top 10 for LLM Applications.

Source: OWASP GenAI Security Project: 2026 Top 10 for LLM Applications announcement · Read 27 Sep 2026

Human in the loop #

A person who can approve or refuse an AI action before it happens. The MCP tools specification says there should always be a human in the loop with the ability to deny tool invocations.

Source: MCP specification: tools (2026-07-28) · Read 27 Sep 2026

Indirect prompt injection #

Instructions hidden in content an AI agent reads, such as a web page, a document or a ticket, that steer it into actions its user did not ask for. Bay's Ghostjacking write-up describes the pattern against trusted agents.

Source: Bay blog: Ghostjacking, an AI agent attack (August 2026) · Read 27 Sep 2026

Local MCP server #

An MCP server that runs on the user's own computer, started by the AI application. The MCP security best practices note that such servers run with the same privileges as the client.

Source: MCP Security Best Practices (2026-07-28) · Read 27 Sep 2026

Managed device #

A computer enrolled in the company's device management or EDR, so security software and settings can be installed and enforced on it.

Managed settings #

Settings for an AI agent that an administrator sets centrally and the user cannot override, such as permission rules, hooks and allowed MCP servers. Bay describes managed settings for Claude Code, Codex and Claude Desktop.

Source: Bay blog: Ghostjacking, an AI agent attack (August 2026) · Read 27 Sep 2026

Masking #

Replacing the sensitive part of a prompt or upload, such as a customer number, before it reaches the AI tool, instead of blocking the whole request. Onyx Security lists mask among its actions.

Source: Onyx Security: AI security platform · Read 27 Sep 2026

MCP (Model Context Protocol) #

An open standard for connecting AI applications to outside systems. An MCP server offers tools and data; an MCP client inside the AI application calls them.

Source: Model Context Protocol: introduction · Read 27 Sep 2026

MCP gateway #

A proxy between MCP clients and MCP servers that can see, filter, block, change or log tool calls. It sees only calls routed through it. See the gateway guide.

MDM (device management) #

Tools such as Intune, JAMF, Kandji or Group Policy that install software and enforce settings on company devices. Harmonic Security names these four for its rollout.

Source: Harmonic Security homepage · Read 27 Sep 2026

On-device control #

An AI control that acts on the laptop itself, through its own agent, a binary pushed through EDR or device management, or hooks inside the AI agent. See the on-device guide.

Policy-only control #

Controlling AI with a written policy, an approved-tools list and the settings tools already offer, without a dedicated product. See policy only.

Prompt injection #

Input written to override an AI model's instructions. Direct injection comes from the user; indirect injection hides in content the model or agent reads.

Secure web gateway (SWG) #

A proxy that inspects and filters employees' web traffic. It is one of the services inside a security service edge.

Security service edge (SSE) #

A cloud service that delivers web filtering, access control and data protection for traffic leaving users' devices. See the network or SSE proxy guide.

Shadow AI #

AI tools, agents, extensions or MCP servers that employees use without the company's approval or knowledge. Prompt Security (SentinelOne) describes inventorying unsanctioned shadow AI.

Source: SentinelOne: Prompt Security · Read 27 Sep 2026

Shadow MCP #

MCP servers that employees connect to their agents without approval, often by adding a line to a configuration file.

Source: SentinelOne: Prompt Security · Read 27 Sep 2026

Simulation mode #

Running a new rule without enforcing it, to see what it would have allowed, asked about or blocked. Bay documents a Simulation Mode.

Source: Bay blog: Ghostjacking, an AI agent attack (August 2026) · Read 27 Sep 2026

stdio transport #

The way a local MCP server talks to the AI application on the same computer, through standard input and output rather than the network. Traffic sent this way never passes a network proxy or gateway.

Source: MCP Security Best Practices (2026-07-28) · Read 27 Sep 2026

Streamable HTTP #

The MCP transport for remote servers, over HTTP. The 2026-07-28 specification added Mcp-Method and Mcp-Name headers to these requests so intermediaries can route them without reading the body.

Source: MCP blog: the 2026-07-28 specification · Read 27 Sep 2026

Tool call #

One request from an AI agent to use a tool, such as running a command or calling an MCP server's tool with arguments. It is the unit most agent rules check.

Source: MCP specification: tools (2026-07-28) · Read 27 Sep 2026

Unmanaged device #

A personal, contractor or other device the company does not enroll in device management or EDR, so it cannot install or enforce security software on it.

Virtual desktop (VDI) #

A desktop that runs on a server and is shown on the user's screen remotely. AI tools used inside it run on the server, where the company controls the network path.