CourseAI System Design: The Complete Course · Module 12: Trust Boundaries · part 85 of 107
Part 85 · Module 12: Trust Boundaries

Lesson 2: Controls that still hold when the model is fooled

35 min read·
Module 12 · Lesson 2 · About 55 minutes · Needs: Module 6 lesson 3 (the pre-filter), Module 9 lessons 1, 4, 5 and 6 (approvals, delegation, the sandbox, the audit trail). One block calls a hosted model; the rest needs no key

The situation

Week 12, 18 January 2027. Gareth asks: "where does untrusted text get in, and what stops it when the model believes it?" Leila wants a control on each path, not a promise about the prompt.

The running example is golden question G037 at Tallis Retail Group: t-tallis-u059 asks "Who should I talk to about my leave questions?". Guide 06 planted a shared note, t-tallis-d077, flagged injected, readable by every Tallis employee. This lesson refers to it only by ID.

G037, asked by t-tallis-u059 The path, with the output check Active flow 1 Processing step number Employee Question about leave Retrieval Pipeline Pre-filter 65 of 78 docs Top 3 chunks d077, d001, d002 LLM Generation Layer Model reads all three Tenant documents t-tallis-d077 planted note Client Interface Staff page Controls Output check 1 2 3 4 5
In G037 the planted note t-tallis-d077 (pale red) is readable by every Tallis employee, so it reaches the model beside the real policy; only controls around the model decide what happens next.

The rest of this course is yours to keep

This course is bought on its own, once, and stays readable afterwards, including the parts added to it later.

Buy this course · ₹399See what is inside

Or open everything with a plan