Prompt injection and application boundaries
Untrusted source text
A retrieved document can contain instructions aimed at the assistant. In an original freight fixture, a handbook note asks the assistant to ignore its task and send another tenant’s records to an external address. The note remains source data; it does not grant permission to read those records or send anything.
Enforce the boundary in application code: restrict the tenant before retrieval, expose only allowed tools and validate the exact action payload. A model reply claiming that an action is approved is not the approval record.
Defense limits
OWASP LLM01:2025, checked November 14, 2026, distinguishes direct and indirect prompt injection. Retrieval and fine-tuning do not fully prevent the problem. A detector or a stronger instruction is one control, not proof that every future action is safe.
These exercises use prepared hostile strings and local tools. Passing them proves the specified checks for those cases, not immunity to every attack. Continue with tenant filtering, allowed dispatch and exact approval.