AI Agent Security
Agents turn a large language model from something that talks into something that acts — and the risk spreads from the text layer into tool calls and permissions. A successful prompt injection in a chat window produces, at worst, a wrong sentence. In an agent, the same injection can become a real file write, a database operation, or a message sent to the outside world.
This section covers how to draw an agent’s permission boundary, why excessive agency happens, how trust is established between collaborating agents, and what the A2A protocol adds to the security picture.
This section currently holds the articles below, in suggested reading order.