How to Detect AI Agents: Types, Examples and Why It Matters

What is an AI agent?

An AI agent is software that uses an AI model to complete tasks for a person. An agent can browse websites, fill in forms, log in to accounts, buy products and make payments. Some agents operate in the person's own browser. Other agents operate in the cloud or send requests directly to APIs.

A small number of agents identify themselves, for example with a signed request. Most agents do not identify themselves. Thus, a business frequently cannot know if a session comes from a person or from an agent.

Why is it important to detect AI agents?

It is important to detect AI agents because a business must know who or what performs each action. Many agents act on behalf of real customers. Thus, a business must not block all agents. But a business must not treat all agents as people. The correct procedure is: identify the agent, then apply the correct policy.

• Most agents do not identify themselves. Signed requests and known identifiers apply to a small part of agent traffic. The business must recognize all other agents.
• An agent that identifies itself still needs checks. The business must know if the person behind the agent is the account owner. The business must also know what the agent will do next.
• Some agents look like people. Some agents are built to look human. Thus, one label on a session is not sufficient.
• Different agents need different policies. A shopping assistant, a cloud service and a script that attacks a login page are different risks.
• Errors cause loss in two directions. If detection misses an agent, fraud can occur. If detection blocks a person, the business loses a customer. Fast typists, autofill and remote desktops can look like machines.

Darwinium puts each session into one of six tiers: agent declared, agent identified, agent inferred, bot driven, crawler or human. Fraud and risk teams use the tier to make a decision.

What are the main types of AI agents?

There are four main types of AI agent: in-browser agents, browser extension agents, cloud agents and agent SDKs. Each type looks legitimate in some ways. Thus, each type is easy to miss.

1. In-browser agents

An in-browser agent is an AI agent that is built into a person's browser or controls that browser. The agent uses the person's real device and home network. Thus, many standard checks see a real customer. But software, not the person, performs the actions.

2. Browser extension agents

A browser extension agent is an extension that clicks and types in a normal browser. The device is real. Frequently, the device shows no risk signals. But the extension, not the person, performs the actions on the page.

3. Cloud agents

A cloud agent is a hosted service that operates a browser in the cloud. The person makes a request in an app or a chat. The cloud agent then completes the task. The browser can look real. But the session does not come from the customer's device.

4. Agent SDKs and frameworks

An agent SDK is a developer toolkit that uses an AI model to control a browser. Developers run agent SDKs on a local computer or on hosted browser infrastructure. Agent SDKs are used for legitimate automation and for large-scale attacks. Agent SDKs show the most machine-like behavior of the four types.

Which AI agents does Darwinium detect?

Darwinium detects AI agents of all four types. The table shows examples for each type.

New agents are added when they are released. If an agent does not match a known name, Darwinium still classifies the session as an agent.

Why detect AI agents at the edge or CDN?

Detection at the edge gives full visibility of AI agents across both applications and APIs. The edge is the CDN layer between users and your servers. All traffic goes through the edge: web pages, mobile apps and APIs.

• Agents do not only use web pages. Many agents send requests directly to APIs. Some agents use MCP servers or headless browsers. These agents frequently do not run page scripts. Detection that operates only in the page cannot see this traffic.
• One view of the full journey. An agent can start in the application and continue through the API. Edge detection connects the two parts into one session. Thus, risk teams see the full sequence of actions.
• Action before the request reaches your servers. At the edge, a business can allow, challenge, rate-limit or block a request. The decision occurs before the request gets to the application or the API.
• Same coverage for all channels. Web, mobile and API traffic get the same agent classification. Thus, an agent cannot avoid detection by changing from the website to the API.
• Less integration work. Edge deployment operates on the CDN that the business already uses. Each application or API does not need separate changes.

Darwinium operates at the edge. Thus, Darwinium gives one consistent view of agent activity across applications and APIs.

Frequently Asked Questions

What is AI agent detection?

AI agent detection is the identification of sessions and requests that AI agents perform. It tells a business if a person, an AI agent, a bot or a crawler performs an action.

Do all AI agents identify themselves?

No. Some agents send a signed request or a known identifier. Most agents do not identify themselves. A business must detect these agents by other methods.

Should a business block all AI agents?

No. Many AI agents act for real customers. The business must identify the agent and then apply the correct policy: allow, challenge, rate-limit or block.

What is the difference between an AI agent and a bot?

A bot follows a fixed script and uses the same path each time. An AI agent uses an AI model to make decisions and can change its actions during a task.

Why is edge detection better than detection only in the page?

Edge detection sees all traffic, including API requests and agents that do not run page scripts. Detection only in the page cannot see these requests.

Next Steps

AI agents now use websites and APIs every day. A business must identify agents to set the correct policy for each one. To see which AI agents use your applications and APIs, contact Darwinium for a demo.