AI Agent Security (Course)
Context Window
Also: context
The limited amount of text a language model can see at once when it decides its next output.
In more detail
Everything the model is given for a turn sits in its context window: the instructions, the user's request, earlier tool results and any files or web pages the harness included. The model has no separate channel marking which parts are trusted, so text from a stranger and text from the user arrive as one stream. That is why untrusted content in the context can steer an agent, and why long sessions that summarise or compact their context create new places for things to go wrong.