Diagrams

Reference diagram showing how file-based AI work can move from upload to extraction, splitting, retrieval, active context, answer generation, and verification checks.

Summary schematic of the request-assembly boundary highlighting common failure modes: instruction injection, precedence confusion, truncation, provenance loss, and memory poisoning.

Reference diagram showing two common failures: chat-derived context influencing effective privileges, and model output persistence/replay into later context; highlights the server-side AuthZ + authoritative state invariant.

Request assembly/context selection schematic with annotated risk points across tools, memory, retrieval/caching, and observability.

Reference model of provenance boundary failure in prompt assembly: authority confusion and cross-turn drift.

Reference schematic of prompt-injection risk in web tool access: routing manipulation and downstream failure modes (content-based instructions, source poisoning, query leakage, unbounded consumption, improper output handling).

Schematic comparing basic-mode hinting vs connector/agent-mode routing to external services, with control/risk callouts.