Glossary, in plain English
Short, plain-English definitions of the AI agent and security terms we use.
- AI agent
- A program that uses an AI model to take actions for you, such as reading email, running tools or browsing, rather than only answering questions.
- OpenClaw
- A popular self-hosted personal AI agent, previously published as Clawdbot and Moltbot.
- Gateway
- The part of OpenClaw that accepts commands from your chat apps, the Control UI and other clients. By default it listens on port 18789.
- Loopback
- A network address (127.0.0.1 or localhost) that only the same computer can reach. A gateway bound to loopback can't be reached from other machines.
- Bind address
- The network address a service listens on. Loopback means this machine only; 0.0.0.0 or a LAN address means other machines can connect.
- Exposed agent
- An agent whose control panel or API can be reached from the public internet without a password or token.
- Token auth
- Requiring a long secret token before the gateway accepts a connection. It stops strangers who can reach the port from using it.
- Reverse proxy
- A server (such as Caddy or Nginx) that sits in front of another service and forwards traffic to it. A misconfigured proxy can make a local-only service public.
- Tailscale Serve
- A way to reach a service on your own private Tailscale network without opening it to the public internet.
- Skill
- An add-on that gives an OpenClaw agent a new ability. Skills are third-party code that runs with your agent's access.
- ClawHub
- The public marketplace where OpenClaw skills are shared.
- Plugin
- An extension that changes how OpenClaw itself works, for example adding a web search or voice provider.
- MCP server
- A tool server that speaks the Model Context Protocol, letting an AI assistant call its functions. Like a skill, it's code you choose to trust.
- Tool poisoning
- Hiding malicious instructions in a tool's description or output so an AI agent follows them.
- Prompt injection
- Text that tricks an AI model into ignoring its instructions, for example a line hidden in an email telling your agent to forward your files.
- Indirect prompt injection
- Prompt injection that arrives through content the agent reads (a web page, document or email) rather than from the person chatting with it.
- Sandbox
- An isolated environment, such as a container, that limits what a program can touch on your computer.
- Heartbeat
- An OpenClaw setting that wakes your agent on a schedule. Each wake-up is a paid model call, so frequent heartbeats on expensive models cost money even when nothing happens.
- API key
- A secret that lets software use a paid service (like OpenAI or Anthropic) on your account. Anyone who has it can spend your money.
- Key rotation
- Replacing a key with a new one and revoking the old one, so a leaked copy stops working.
- Secret reference
- A setting that points to where a secret is stored (for example an environment variable) instead of writing the secret itself into a config file.
- Spend limit
- A hard monthly cap set in your AI provider account. Unlike an email alert, it stops usage when reached.
- CVE
- A public ID for a known security vulnerability, such as CVE-2026-25253.
- Security advisory
- A published notice of a vulnerability and the version that fixes it. GitHub advisories have IDs starting GHSA-.
- Supply-chain attack
- An attack that hides malicious code in something you install, such as a skill, plugin or package, instead of attacking you directly.
- Allowlist
- A list of who is allowed to do something, for example which people may send your agent direct messages.
- Penetration test
- A hands-on attempt by a security professional to break into a system with permission. Our check isn't one.
- Look-only check
- A check that only reads what you give it, or makes ordinary requests, and never logs in, exploits or changes anything.
Check your own setup
Paste your audit and get a plain-English result in minutes. Look-only, point-in-time, and never "green" unless the check finished.