Find out how an AI assistant can be tricked, and make sure a tricked assistant still cannot do harm. You list the risks of Larkfield's support assistant, attack it with harmless injections and fix the consequences in code. Then you map its personal data, design human review, and finish with a security test set and an incident plan.
Read every lesson as a guest, or sign in to enroll and track your progress.
Enroll to save the lessons that you complete. It is free.
By the end of this course, you will be able to:
An assistant that reads documents and uses tools does what its text tells it to do. The problem is that some of that text comes from people you do not trust: a customer’s message, a file, a web page. A sentence hidden in any of them can ask the assistant to reveal a secret or to send something where it must not go. This attack is called prompt injection, and a better prompt alone does not stop it.
In this course, you attack Larkfield’s support assistant yourself, with harmless attacks against a practice app on your computer. You measure what a strong prompt and a filter can stop, and what they cannot. Then you add the controls that hold even when the model is fooled: checks in code, tools with the least access they need, allowed destinations and a person’s approval before every change.
Security is not the only job. You also map where customers’ personal data goes and how to delete it, decide when a person must review a decision, and plan what to do when something goes wrong.
You do not need an account or a paid key. Every model decision has a recorded response from a real model, and the website, files and secrets that the attacks target are made up. Diagrams and explorers carry the main ideas, so the text stays short.
You completed Building Reliable Applications with LLM APIs, Build and Deploy Your First API and Authentication, Permissions and Multi-Tenant Applications, or you have the same skills. RAG, agents and evaluation courses help but are not required. You do not need an API key or a paid account. Every model decision has a recorded response, every attack is harmless and runs only against a practice app on your computer, and live calls with a free local model or your own key are optional.
Courses to take first: Building Reliable Applications with LLM APIs, Build and Deploy Your First API, Authentication, Permissions and Multi-Tenant Applications.
People who have built an application on a model API and an API with sign-in and permissions, as in Building Reliable Applications with LLM APIs and Authentication, Permissions and Multi-Tenant Applications, and now must make an assistant with tools safe to run. Developers, data scientists, testers and professionals moving into AI engineering.
You can study the ideas of this course without writing code. Read modules 1, 2, 4 and 5 (about 7–10 hours), answer their knowledge checks, and get a free concepts certificate. You can come back for the practice later.
The paths for people who lead or use AI take this course this way. See the paths
Experienced AI practitioners designed this course. AI helped us draft the content; people chose the teaching approach and the examples, and reviewed every lesson. The outputs that you see are the actual results of running the code. How we make our courses
Map what you must protect before you defend it. List the assistant's assets and actors, draw where trusted instructions meet untrusted content, and rank the paths to an unauthorized read or action.
See prompt injection happen, then limit what it can do. Run harmless attacks through messages, documents, files, tool results and a web page, measure what prompts and a filter stop, and add checks in code that a fooled model cannot pass.
Make every tool safe on its own. Validate each operation, give each role only the tools it needs, allow network requests only to named hosts, and keep secrets and message text out of the model's reach and the logs.
Know where personal data lives and make it go away. Find every copy of a customer's data, keep only what has a purpose, make a deletion reach caches, indexes, logs and backups, and know which questions need a lawyer.
Decide when a person must look. Find who is hurt by a wrong answer, test the assistant by language and customer group, and design a review step where the reviewer checks evidence, not only a button.
Test, watch and respond. Measure attack success next to task success, detect attacks without collecting message content, and plan the response to an incident and the risks that remain.
Two end-to-end case studies. Find and mask personal data in open text data before it reaches a model, and threat-model and test an open-source AI application. Complete at least one for the certificate; each one is a portfolio piece.
Complete at least one case study for the certificate. The others are optional. What each one covers
Use every module in one practice project. You collect evidence, and check it against the course outcomes.
Realistic projects from start to finish, on real data, with every output explained. Each one makes a portfolio piece. Complete at least one for the certificate; the others are optional.
Find and remove personal data before it reaches a model2 hr 30 min
Detect and mask personal data in open text data, measure misses and over-masking by type and language, and decide what still needs a person.
Threat-model and test an open-source AI application4 hr
Run a small open-source AI application locally, draw its data flow and trust boundaries, test it with harmless injection and access cases, and write its residual-risk register.
Loading the reviews…
Took this course? Sign in to review it. Learners can write a review after they complete 25% of the lessons.
