Defending Meridian Assurance's claims assistant, which retrieves over policy documents and can issue refunds, after a poisoned document made it disclose another customer's data: three inputs pass through the eight-layer guardrail stack — a claimant's status question passes every layer and gets only their own claim, a planted instruction in a policy PDF is stopped at input and context isolation, and an injected refund instruction in tool output is stopped at tool authorisation by human approval; canary tokens and provenance logs keep the evidence, and an evidence pack maps the controls to the EU AI Act for the regulator.

CY-EXPERT-017Expert Level · LLM Security

Prompt Injection Defensibility — The Guardrail Stack

An expert-level course that builds and proves the case that an organization's own large-language-model systems cannot be made to take unauthorised actions or disclose data through prompt injection, to a standard that satisfies regulators, auditors, and litigation, across eight guardrail layers from threat modelling to a regulator-ready evidence pack.

18 hours8 chapters32 lessons225 activities8 labs

What you'll learn

Name direct and indirect prompt injection on a duty-mapped threat model owned and reviewed on a defined cadence.
Isolate retrieval corpus ingestion, spotlighting, context budgeting, and tenant boundaries so a planted document cannot act as a command.
Authorise every tool call with least privilege, human-in-the-loop approval, and user-bound capability tokens, closing the confused-deputy gap.
Assemble a regulator- and board-ready evidence pack mapped to AI Act, NIST AI RMF, and ISO/IEC 42001, with defensibility metrics.

Price:৳12,000