Context window
The text a model can take into account at once, measured in tokens: system prompt, conversation, tool definitions, file contents and tool results. For security it is the attack surface, because everything in the window can steer the model whoever wrote it; reading a file or a web page into the window is the moment an indirect prompt injection lands.
See also indirect prompt injection · tool registration · context tree
Related terms
-
Indirect prompt injection
Attack concepts
The instructions are planted in content the agent reads rather than text the user types — a web page, a GitHub issue, a code comment, a dependency's…
-
Tool registration
Agents & architecture
The handshake in which a server tells the host what tools it offers, and the host puts those names, descriptions and schemas into the model's context.
-
Context tree
Defence & response
In AI incident forensics, a reconstruction of everything that entered a model's context during a session and where each piece came from — prompts, files…
Agents & architecture
Where the attacks above actually land. The vocabulary here is worth precision, because the difference between "the model decided" and "the tool ran" is the difference between a risk and an incident.