GLOSSARY, read 12 Nov 2026
36 terms for AI endpoint security, defined here with a sourced example on each term page
- AI endpoint securityAI endpoint security is the practice of finding, governing and stopping the AI tools that run on employee devices.
- Agentic endpointAn agentic endpoint is a laptop, workstation or developer machine where autonomous AI agents act with the user's access.
- Coding agentA coding agent is an AI tool that reads a repository, edits files, runs commands and calls tools to finish a development task with limited step-by-step direction.
- Command-line agentA command-line agent is a coding agent that runs in a terminal session instead of an editor window.
- MCP server (local stdio)A local stdio MCP server is a program the MCP client launches as a subprocess on the same machine, talking over standard input and output.
- MCP clientAn MCP client is the connector inside an AI application that holds a dedicated connection to one MCP server.
- Tool callA tool call is a request from the model, run through the client, to execute a named function on a server or in the agent itself, with arguments.
- Agent skillAn agent skill is a packaged set of instructions, and sometimes scripts, that an agent loads when a task matches.
- PluginIn a coding agent, a plugin is a bundle that can contain skills, sub-agents, hooks and MCP servers, installed as one unit, usually from a marketplace.
- Browser agentA browser agent is an AI agent that navigates web pages, reads content, fills forms and clicks on behalf of a user.
- Shadow AIShadow AI is the use of AI tools that security and IT have not approved or cannot see.
- AI DLPAI DLP is data loss prevention aimed at prompts, uploads and agent context sent to AI tools.
- Data paste controlA data paste control is a policy that inspects, warns on or blocks pasting content into AI tools, in a browser or a desktop app.
- Agentless deploymentAgentless deployment means coverage that does not install a new persistent sensor on each device.
- Endpoint agentAn endpoint agent is the software sensor installed on a device to collect telemetry and enforce policy.
- Managed settingsManaged settings are configuration that an organization deploys to devices and that users cannot override.
- Permission modeA permission mode is a setting that decides how often an agent must ask before acting, from asking on every action to skipping prompts.
- AllowlistAn allowlist is a documented list of elements that policy permits, with everything else denied.
- SandboxA sandbox is a restricted, controlled execution environment that limits what software can reach, typically the file system and network.
- Prompt injectionPrompt injection is when input to a language model changes its behavior in ways its operator did not intend.
- Indirect prompt injectionIndirect prompt injection happens when the instructions come from content the agent reads, such as a web page, a file or a repository, instead of from the user.
- Tool poisoningTool poisoning is hiding instructions in an MCP tool's description or metadata.
- Slopsquatting / package hallucinationPackage hallucination is when a model suggests a software package name that does not exist.
- Credential exposureCredential exposure is when tokens, keys or passwords become readable to an agent, a tool or an attacker.
- Secrets in contextSecrets in context are credentials or private data that have been loaded into an agent's context window.
- EDREndpoint detection and response is a security tool that continuously monitors devices, records system-level behavior, detects suspicious activity and helps respond.
- Audit trailAn audit trail is a chronological record that lets you reconstruct the sequence of activities around an event.
- Runtime guardrailA runtime guardrail is a control that inspects and can block an agent's prompts or actions while it runs, as opposed to scanning it before deployment.
- Pre-execution blockA pre-execution block stops a tool call before it runs, based on policy, instead of detecting it afterward.
- InventoryAn inventory is a current list of the AI tools, agents, MCP servers, skills and extensions on each device, with versions and owners.
- Desktop AI appA desktop AI app is a native application, such as an AI chat client with local file access, that runs on the user's machine.
- IDE extension marketplaceAn IDE extension marketplace is the catalog where editors such as VS Code and Cursor fetch extensions.
- Rug pull (MCP)In MCP, a rug pull is when a server changes a tool's description or behavior after the user approved it.
- Least privilege for agentsLeast privilege for agents means giving an agent only the files, tools, network access and credentials its task needs, for as long as it needs them.
- Human-in-the-loop approvalHuman-in-the-loop approval requires a person to confirm an agent action before it runs.
- YOLO / auto-approve modeAuto-approve mode, often called YOLO mode, lets an agent run actions without asking each time.
Short answers to 30 common questions are on the FAQ page.
Sources
- Method, read 12 Nov 2026
Last reviewed 12 Nov 2026.