AI agents

Agents are essentially large language models (LLMs) that don’t just generate responses, they can also perform actions. In a simple setup, you might give an LLM access to tools like:

• Internet search
• The ability to write and run code
• A calculator

When an LLM can actively use these tools, it becomes what we call an agent.

More generally, AI agents can interact with tools, websites, and systems to complete tasks such as:

• Filling out forms
• Collecting and processing data
• Automating workflows

This ability to act turns an LLM from something that only gives one-time answers into a system that can carry out ongoing tasks and move step by step toward a goal without needing constant human input.

Why do agents matter?

Because they can take action, agents open up powerful new possibilities, including:

• Embodied LLMs connected to real world devices such as robots, allowing them to perform physical actions
• Autonomous systems that can operate continuously without human input
• Agents that interact with external APIs, databases, and services on your behalf

However, this also raises the stakes. If an attacker manipulates such an agent, it could lead to real world consequences. For instance, imagine a humanoid robot powered by an LLM navigating a public space. If provoked, how would it respond? The behavior isn’t always predictable.

This is why AI red teaming is so important. By testing these systems for weaknesses, we can:

• Identify vulnerabilities before they are exploited
• Understand how agents behave under adversarial conditions
• Design agents with stronger safety measures, reducing the chances of harmful outcomes