What do AI coding security words mean? A plain glossary

AI coding security has a lot of jargon. This glossary explains more than 20 common terms in plain words, grouped by topic: the AI itself, the software it touches, and the controls that keep it safe. Read it once, keep it handy, and the next meeting will make a lot more sense.

TL;DR: Three groups of words cover most conversations. AI words (model, prompt, agent, MCP) describe the tool. Software words (dependency, SBOM, CVE, reachability) describe what it changes. Control words (sandbox, permission, gate, pull request) describe how you keep it in check.

Definition: glossary. A glossary is a short list of terms with simple explanations. This one is written for smart readers who do not work in security every day.

What do the AI words mean?

Model. The trained AI system that produces answers and code. Companies such as Anthropic and OpenAI build models and offer them through products. The model is the engine. The product is the car around it.

Prompt. The instruction you give the AI, such as "write a function that sorts a list." Better prompts usually give better results. Prompts can also contain hidden text, which matters for the risk called prompt injection, defined below.

Context window. The amount of text the AI can keep in mind at one time. If a conversation or project is too long, older parts may drop out. That is why an assistant sometimes forgets a rule you gave it earlier.

Hallucination. When an AI states something that is not true, such as naming a software package that does not exist. It is not lying on purpose. It generates text that sounds right, and sometimes the text is wrong.

Coding assistant. An AI tool that suggests or explains code. You decide what to copy into your project.

Coding agent. An AI tool that can also act: open files, edit them, run commands and decide the next step. Anthropic's documentation describes Claude Code as an agentic coding tool that reads your codebase, edits files, runs commands and connects to your development tools. Our post on what an AI agent is and how it differs from a chatbot goes deeper.

Tool call. A step where the AI uses a tool, such as running a test, reading a file or asking a database a question. Each tool call is an action that can be limited, approved or logged.

MCP (Model Context Protocol). The MCP project calls it an open standard for connecting AI applications to external systems. Think of a standard plug that lets an assistant reach a calendar, a database or a security tool. Our glossary of MCP security terms covers the details.

Prompt injection. A trick in which text the AI reads, such as a web page, a file or an issue comment, contains instructions that try to steer it. The AI may follow the hidden instruction instead of yours. It is one reason agents should have limited access.

What do the software words mean?

Dependency. A library of code written by someone else that your project uses. Most software is mostly dependencies. See what a dependency is in programming.

Package manager. A tool that downloads and installs dependencies by name, such as npm or pip. See what a package manager is.

Lockfile. A file that records the exact version of each dependency so every install is the same. AI tools sometimes regenerate it, which can change many versions at once.

Vulnerability. A flaw in software that an attacker might be able to misuse.

CVE. A public identifier for one known vulnerability, such as CVE-2021-23337. It lets everyone talk about the same flaw. See how to read a CVE.

SBOM (software bill of materials). A list of every component inside an application, like the ingredients list on food packaging.

Reachability. Whether your code can actually reach the risky part of a library. A flaw in a function you never call is usually lower priority than one in a function you use every day.

Slopsquatting. An attack in which someone registers a package name that AI tools tend to invent, hoping developers install it. Our post on how slopsquatting exploits AI-hallucinated package names explains it.

What do the control words mean?

Sandbox. A fenced-off area where a program has limited access, such as one folder and no internet.

Permission mode or approval policy. A setting that decides when a tool must ask you before acting. OpenAI's Codex documentation separates the sandbox, which sets what is technically possible, from the approval policy, which sets when it must ask.

Least privilege. Give a tool, person or agent only the access it needs for the job and nothing more.

Pull request. A proposed code change that people and tools can check before it is accepted. See what a pull request is.

Gate. An automatic rule that blocks a change if a check fails, such as "no new critical flaws with a fix available."

How do the words fit together in one table?

What does a worked example look like?

The sentence below might come from a team update. All details are invented for illustration.

"Our coding agent opened a pull request that added a new dependency. The scan found a CVE, but reachability says we do not call the affected function, so the gate passed with a warning."

Here is the same sentence in plain words, one step at a time.

1. "Our coding agent opened a pull request." An AI tool that can act proposed a code change for people to check.
2. "Added a new dependency." The change brought in a library someone else wrote.
3. "The scan found a CVE." A tool matched that library against a list of known flaws and found one with a public ID.
4. "Reachability says we do not call the affected function." Our code never uses the risky part, so the flaw is less urgent.
5. "The gate passed with a warning." The automatic rule allowed the change but flagged it for a person to review.

Now you can ask good questions. Was the dependency needed? Is there a fixed version? Who decided that a warning was acceptable, and where is that recorded?

How do you use this glossary with your team?

Try this short routine in your next meeting about AI tools.

1. Pick three terms from the table that the team uses loosely.
2. Ask each person to define them in one sentence.
3. Compare. Where definitions differ, agree on one and write it down.
4. Add the agreed list to your onboarding notes.

Shared words prevent a surprising number of misunderstandings, especially between engineers, managers and security staff.

Where does Safeguard fit in?

Safeguard covers several of the software and control words above. It builds an SBOM, scans dependencies for known flaws and CVEs, and uses reachability analysis to show which ones your code can trigger. Policies and gates can block a pull request that breaks your rules. For the AI side, Safeguard offers an MCP Guardrail to control what connected AI tools may do, and agent identity to track which agent did what.

Fix pull requests are available but off by default, with every category on Manual. The Safeguard CLI runs on your own machine, sends only scan results (not your source code) to the platform, and can also write code. Installers are listed on the download page, and you can try the free scan at gold.safeguard.sh/scan.

Try it yourself

Open the demo below and walk through connecting an AI tool to an MCP server. As you click, match each step to a word in this glossary, such as tool call, permission and least privilege.

Click through it at your own pace, or open it full screen.

Quick check

Sources

• Claude Code overview, Anthropic documentation
• Codex agent approvals and security, OpenAI developer documentation
• What is the Model Context Protocol (MCP)?
• Safeguard downloads
• Safeguard CLI

Last reviewed: 19 October 2026.