# What is AI agent security?

AI agent security is protecting systems where a model can act (call tools, run code, make payments, direct other agents), so that neither an attacker nor the model's own mistake turns an action into harm.

## Why agents are different

A chatbot writes text, and a person decides what to do with it. An agent's output is an action that runs with nobody in between. The agent also reads content it did not write (web pages, emails, files, tool results, messages from other agents), and any of it may carry instructions.

## Main risks

- **Prompt injection.** Text in something the agent reads tells it what to do. See [indirect prompt injection](/glossary/indirect-prompt-injection).
- **[Excessive agency](/glossary/excessive-agency).** The agent has more tools, permissions or autonomy than its task needs, so one wrong step does real damage.
- **[Data exfiltration](/glossary/data-exfiltration).** The agent sends private data out through a URL, an email or a tool.

## Controls

- Give each agent only the tools and access its task needs.
- Check every tool call against policy before it runs. See [guardrails for tool calls](/glossary/tool-call-guardrails).
- Scan tool results and retrieved content before the model acts on them.
- Hold actions that cannot be undone for a person's approval.
- Record every step, so an incident can be reconstructed.

[How to make AI agents safe](/glossary/how-to-make-ai-agents-safe) goes through each.

## With White Circle

[Control](/products/control) checks tool calls (terminal commands, code execution, API calls) before they run, and blocks injected instructions before they reach a tool.

