Web tool access — prompt injection risk map

Reference diagram of prompt-injection risks in web tool access, including routing manipulation and downstream failure modes.

Diagram showing Client UI to Core LLM to Tool Router to Web tool within an LLM boundary. A central note states prompt injection influences routing and follow-up actions. On the right, risk boxes list content-based instructions, SEO/source poisoning, query leakage (PII/secrets), unbounded consumption, and improper output handling.
Prompt injection risks in web tool access (reference model).

Overview

A vendor-agnostic risk map for systems where an LLM can route to a web browsing/retrieval tool.

Text alternative (long description)

  • A large box labeled LLM boundary contains Core LLM and Tool Router.
  • Flow: Client UI → Core LLM → Tool Router → Web tool.
  • A callout states: Prompt injection influences routing + follow-up actions.
  • Risk list (right side): Content-based instructions, SEO / source poisoning, Query leakage (PII / secrets), Unbounded consumption, Improper output handling.

Scope and limitations

  • This diagram is a reference checklist; it does not assert any specific product implementation.
  • “Web tool” refers to any browsing/retrieval connector that fetches untrusted external content.