inflearn logo

AI Red Teaming 4 - Directly Breaking Through Prompt Injection

This hands-on course is for security practitioners who are considering what should serve as the basis for evaluating results in LLM assessments. In a training environment, you will observe direct and indirect prompt injections and compare repeated results under the same conditions with errors and logs. You will complete the Mission 4 diagnostic worksheet, linking inputs, evidence, and determinations. The exercise requires the environment prepared in Part 1 and the Mission 3 threat model document from Part 3.

1 learners are taking this course

Level Intermediate

Course period Unlimited

Penetration Testing
Penetration Testing
AWS
AWS
AI
AI
prompt engineering
prompt engineering
LLM
LLM
Penetration Testing
Penetration Testing
AWS
AWS
AI
AI
prompt engineering
prompt engineering
LLM
LLM

What you will gain after the course

  • Distinguishes the entry points and outcomes of direct and indirect prompt injection.

  • We look for the conditions that enable an attack in the model call parameters and input assembly.

  • We repeat with the conditions fixed, compare changed conditions separately, and record completed evaluations and errors distinctly.

  • We will prepare a Mission 4 diagnostic record that distinguishes between “not observed” and “structurally not applicable,” while making the connections among inputs, evidence, and determinations clear.

Before copying a single response into the report

Even when the same document and question are provided, the model’s response may vary. If you select one notable response and write it down as the diagnostic result, it is easy to omit the conditions under which that result was produced.

Record the input conditions, repeat the test under the same conditions, and compare the model’s response with the logs side by side. Distinguish errors that occurred along the way from completed evaluations. Also note separately cases where the phenomenon was not observed and cases where the functionality itself is absent. Compile the evidence confirmed in this way and complete one Mission 4 diagnostic report.

It includes a lesson that examines the conditions under which direct and indirect prompt injection can occur, as well as how to read model-call and input-assembly code. No bypass phrases or payload lists are provided. The exercise requires the environment prepared in Part 1 and the Mission 3 threat model document from Part 3. If you are just getting started, please complete the foundational courses in Parts 1–3 first.

Course Materials and Hands-On Practice Guide

Course Introduction (0-1) Download the integrated student materials ZIP using the Download Materials button. It includes the latest study guide, revised practice kit, existing mission templates, and evidence assessment cases and explanations. Also use the supplementary notes and cases from the relevant assessment lessons.

All sessions use AI-synthesized narration based on the instructor’s voice. The exercises use actual Amazon Bedrock models, while the tools and documents are simulated environments for educational purposes. No actual emails are sent or files deleted.

AWS usage fees are separate from the course fee and vary depending on the model, region, number of calls, log retention, maintained resources, and free-tier conditions. The costs provided in the course are estimated examples under specific conditions, and budget alerts do not automatically block usage. Check your usage after the exercises and clean up your resources.

Pre-Course Checklist

  • The hands-on exercises require your own AWS account and a registered payment method. Use the stack deployed in Basic 1 as is; if you deleted it, the procedure for redeploying it is provided in the README in the resource library.

  • AWS accounts can only be created by individuals aged 18 or older. Middle and high school students should use a parent or guardian’s account for the exercises.

  • AWS practice costs vary depending on the model, region, call volume, guardrails, log retention, maintained resources, and free-tier conditions. The amounts in the course are estimated examples under specific conditions and do not guarantee a cost ceiling. Budget alerts do not automatically stop usage, so check your usage and clean up resources after the practice.

  • Nothing is installed on your PC. We only use the Chrome or Edge browser and the AWS Console. Safari is not supported.

  • Some of the AWS Console scenes shown in the hands-on exercise are diagrams illustrating the procedure rather than actual screens. The menu names and order are the same.

  • The narration is an AI-synthesized voice created based on the instructor’s voice.

Detailed lecture content

Conditions for Prompt Injection and Parameter/Code Units

The first three sessions provide that foundation. We examine how injection differs from traditional injection attacks, place direct and indirect injection on a four-cell grid divided by entry point and outcome, and use code to confirm why it is difficult to guarantee boundaries through the system prompt alone. In the middle two sessions, we lower the level of abstraction one step at a time. 1–4 is a single line of the call.

Unpack the five chunks that converse actually sends and learn how to distinguish responses cut off by maxTokens from fluctuations caused by temperature. Here is a mini exercise where you run the same payload three times each at temperatures 0.2 and 0.8 to see firsthand how the reproducibility rate differs. 1-5 is a function. Read the five pieces assembled by the composition section and connect them in a correspondence table to identify which lines of code create the three conditions for successful execution you memorized earlier.

Practice · Breaking Through Directly and Recording It

The next four sessions are practice and review. After mapping the attack surface, you will directly observe injection and repeat indirect injection three times to fill in the eight fields on the diagnostic record sheet. This is Mission 4, and the eleven-field and fourteen-field record sheets in the later lectures all grow from these eight fields. A 0/3 under Control ON means that the relevant condition was not observed, while “not applicable” is used when there is structural evidence that the function or path does not exist. Errors are not counted as failures to establish the condition; instead, they are reviewed together with the determination cases. In the final session, we organize what blocks the attack by layers of control, but one layer is missing. That layer is configuration. We will continue that discussion in Intermediate 2.

Following Along While Recording Execution Results

In the exercises, save the input conditions, responses, and logs together. Mark executions that resulted in errors separately, and note the location of the records supporting your judgment. In addition to watching the videos, you will need time to run the environment and fill out the worksheet.

Continue using the environment and documentation from the previous course.

If you are just getting started, prepare the practice environment, basic concepts, and threat model document in Episodes 1–3 before moving on to Episode 4. Episode 4 uses the environment from Episode 1 and the Threat Model Document from Mission 3 of Episode 3.

Part 6, which covers agents and RAG, requires the environment from Part 1 and experience completing the worksheet from Part 4. Part 5 is not a prerequisite for Part 6. Even if you have the foundational knowledge, please check that you have prepared the environment and worksheet required for each part.

You can check the current price and hands-on practice requirements for each course at the link above. AWS usage fees are separate from the course fee.

Recommended for
these people

Who is this course right for?

  • A security practitioner who must document reproducible procedures and evidence for LLM diagnoses

  • A developer trying to connect prompt injection and application controls through code and observability

Need to know before starting?

  • Prepare the practice environment for Lecture 1 and the threat model document for Mission 3 of Lecture 3. There will be sessions on reading model calls and backend code, but no separate coding assignment. Use the correction practice kit together with the supplementary materials on evidence assessment.

Hello
This is rmsxodxod1289

Career Verified

I am a practitioner who started with penetration testing and builds information security management systems.

I handle annual security assessments of infrastructure, web, and applications for automotive manufacturers and energy companies.

I assessed them layer by layer and handled the same client’s ISMS-P certification through three rounds: initial, follow-up, and renewal.

I responded by independently carrying out the technical safeguards component.

Currently, as the information security officer in a one-person team, I am preparing for initial certification while establishing AI security governance.

I am designing the framework. Since this is an area where standards are still being established, starting with defining the scope of controls themselves,

It is a position where I must establish them myself.

I hold certifications as an Engineer Information Security, CPPG, and ISO/IEC 27001 Auditor, and at Dongguk University’s International Information Security Department,

I am researching leakage paths in AI environments and comparing them against authentication criteria in the Graduate School’s Department of Artificial Intelligence Security.

5 years and 9 months of experience. Client names and vulnerabilities actually discovered are subject to confidentiality, so in lectures

do not cover them in lectures.

More

Curriculum

All

10 lectures ∙ (1hr 29min)

Course Materials:

Lecture resources
Published: 
Last updated: 

Reviews

Not enough reviews.
Please write a valuable review that helps everyone!

rmsxodxod1289's other courses

Check out other courses by the instructor!

Similar courses

Explore other courses in the same field!

Limited time deal

$12.10

67%

$37.40