Practical Security for LLM Agents

As AI rapidly permeates daily life and industries, attack techniques targeting these systems are evolving alongside them. This course is a practical, hands-on program designed to help you understand the latest security threats surrounding Generative AI and LLMs, automate security tasks using AI, and experience LLM jailbreak attacks and defenses firsthand.

We examine AI security incidents and the role of red teams, then set up an LLM practice environment in our own AWS account. From capturing the initial response to checking usage and cleaning up resources, we record everything in the Mission 0 checklist. The prompt injection diagnosis exercise continues in Part 4.

The era when AI simply answered questions is over. Now is the age of LLM Agents that make decisions and take action on their own. This beginner-friendly course lets you learn the core principles and structure of agents by implementing them firsthand in just 90 minutes of hands-on practice. With complex theory kept to a minimum and a code-focused approach, you’ll directly experience how AI makes decisions and uses tools. Go beyond prompt engineering and take your first step toward AI automation with us.

Solve the point where 80% get stuck at makeathons. Complete Raspberry Pi · Computer Vision · LLM · 3D Design in 4 weeks! Achieve top rankings at makeathons with a demonstrable smart mirror PoC!

Review the AI service’s settings, including keys, access scope, call limits, execution permissions, guardrails, and logging. Compare the controls before and after while keeping other conditions unchanged, and review output handling to create the Mission 4-2 settings checklist and diagnostic record.

Hashing, salting, JWT, symmetric and asymmetric keys, and even HTTPS and TLS. The backbone of security that we use every day yet always found vague and unsettling. If we have to face the countless neglected threats and technical debt someday, let’s properly tackle them once and for all this time.

As AI systems become agentic and expand beyond text into image, audio, and multimodal capabilities, security threats are becoming significantly more complex. This lecture analyzes the latest security threats and attack techniques faced by AI Agents and Multimodal LLMs, as well as real-world cases, from an AI Red Teaming perspective.

Claude Code handles judgment and execution, while n8n takes care of triggers, connections, and scheduling. By following the sequence of Concept → Connection → Practice, repetitive tasks are designed into a flow of Input → Judgment → Execution → Verification → Delivery. With error handling, retries, and human approval included, the system operates unmanned through scheduled execution. Even non-developers can build their own automation systems without having to learn coding from scratch.

Deploy a Spring Boot SNS app to local Kubernetes, and automate deployments with GitHub Actions and Argo CD. Identify the causes of failures through metrics, logs, and traces, and review them based on AI analysis results.

We invite you to become a ‘Top-Tier Enterprise AI Systems Architect’ Do not remain a micro-level technician (Prompt Writer) who merely assembles libraries created by others or polishes prompt sentences. From mathematical first principles to runtime control-loop kernels and autonomous recovery infrastructure — through 40 years of practical engineering insight that has driven industrial operations in South Korea, I will firmly guide you to become a top-level ‘Enterprise AI Systems Architect’ who protects your company’s business and digital sovereignty. Now, begin the great journey of transforming uncontrollable AI into a perfect industrial core. “Why does an AI that ran perfectly in a demo start spouting strange nonsense or bring the pipeline to a halt as soon as it is deployed to production?” “Why do AI projects that so many companies launch with great enthusiasm ultimately fail to escape the ‘PoC (Proof of Concept) swamp’ and get abandoned?” These are the most desperate questions that countless developers and systems architects I have met in the field have asked me. Hello. I am Hongjip Kim, who spent more than 40 years in the powertrain R&D division of the South Korean automotive industry (including 17 years as an executive in Hyundai Motor Group R&D), led the export of engine and drivetrain technologies worth 130 billion won, and commanded operations as the CEO/CTO of a listed mid-sized manufacturing company. From the perspective of an engineer who has spent decades in mechanical engineering and precision-control environments, a ‘risk of fatal uncertainty’ lurks behind today’s frenzy to adopt generative AI. In mission-critical industrial environments where even a mere 0.1% error or mistake can directly result in the destruction of physical equipment worth hundreds of millions of won and an enterprise-wide shutdown, integrating a probabilistic chatbot that simply relies on probability to generate plausible text directly into a service is like carrying a ticking time bomb into a fire. This course was designed to impart the “practical engineering methodology for training uncontrollable probabilistic AI into the most robust industrial control core, one that allows not even 0.000% of syntax errors.” 1. The 3 Major Fatal Bottlenecks You Face in the Field (Problem) Many engineers attempt to challenge production environments using an approach limited to one-off prompt tuning (Vibe Prompting), only to become frustrated when they encounter the following 3 structural defects: Finite Context and ‘Lost in the Middle’: As conversations or multistep executions grow longer, critical system instructions are pushed aside and the context window is exhausted, causing the system to lose its memory and amplifying hallucinations. Nondeterministic Outputs and Cascading Pipeline Collapse: Tiny probabilistic fluctuations, such as one missing closing bracket in JSON or an incorrect type, can paralyze legacy ERP/CRM API calls and destroy data integrity. Untrusted Code Execution and Infrastructure Security Bankruptcy: When scripts autonomously generated by an agent run in a non-isolated host OS environment, they create risks such as kernel escapes and sensitive-data leakage. As such, AI malfunctions in industrial environments are not merely quantitative accuracy problems; they become a massive financial disaster in the form of asymmetric business loss (Asymmetric Business Loss). 2. ‘How’ to Solve These Problems Perfectly (Solution) This course differs fundamentally from commercial courses that teach only how to use fragmented off-the-shelf tools. By combining 40 years of hands-on field expertise with computer systems engineering, it delivers 3 hardcore architectures for vertically integrated, system-level control of probabilistic AI: ① Move Beyond Prompting to ‘LLM-OS & Context Orchestration’ Following Andrej Karpathy’s paradigm, redefine the LLM not as a conversational respondent but as the central computational core (CPU) of a computer. Systematically organize the context window as ultra-low-latency volatile RAM, the internal vector DB as Disk, and external APIs as an I/O bus. Through MemGPT hierarchical virtual-memory paging, preserve critical state without saturation even during infinite conversations, while scaling a dynamic agent network through MCP-based Liquid Swarm MAS, minimizing token costs. ② ‘Deterministic Harness Engineering’ That Constrains Probabilistic Hallucinations to 0.000% Leave the language model’s creative flexibility open only during the Reasoning phase, while firmly confining the physical output layer that interacts with external systems using mathematical Boolean algebra and compilers. Apply EBNF / XGrammar-based CUDA logit masking to forcibly block the generation probability of invalid tokens at P = 0.000%. Form a physical security isolation ring so that code generated by agents runs only once inside a Firecracker MicroVM micro-virtualization sandbox with 5ms snapshot restoration capability. ③ ‘High-Reliability MLOps·LLMOps Autonomous Recovery Infrastructure’ Without Wasting GPU Resources Build vLLM PagedAttention and S-LoRA hot-swapping to suppress idle VRAM waste to below 4%. When a multi-agent API fails, restore data consistency 100% through SagaLLM reverse compensation transactions. Through an OpenTelemetry/Langfuse distributed observability network, defend system availability by executing an unmanned automatic canary rollback whenever a decline in the RAGAS faithfulness score (Data Drift) is detected. 3. Practical Value That Drives Spiral Growth (Pedagogy & ROI) This course is systematically designed according to the 5 levels of Bloom's Taxonomy. Starting from simple theoretical recall (Remember), learners directly code and validate a 5-stage debugging pipeline (analyzing deceptive baseline trap source code ➔ deconstructing adversarial engineering ➔ establishing a Robust integrity pipeline ➔ calculating the TEI ROI formula ➔ SOTIF/ISO 42001 governance). Going beyond the intuitive satisfaction of simply “working well,” I will apply the Forrester TEI (Total Economic Impact) financial model to visualize, from an executive perspective, how much the AI system you build reduces your company’s rework cost (Rework Cost), avoids failure risks worth hundreds of millions of won, and proves an ROI of more than 200%.

Simple tutorials alone make it difficult to apply in practice. I will clearly pass on my professional know-how, covering complex state management and multi-agent design methods.

Information security is difficult and too technical? No. Start learning the basics of information security to create a secure web right now!

Django Base Camp is a course where beginners to intermediate learners can systematically master the core of the Django web framework. Through project-oriented practical lessons, you can quickly develop the ability to build web applications.

We explore how to create an AI judge by combining knowledge graphs, reasoning systems, and various agent frameworks such as CrewAI, LangChain, and AutoGen, which enable the implementation of LLM-based multi-agent technology.

Build a Saju (Four Pillars of Destiny) analysis AI agent that integrates Manseryeok, Astrology, Zi Wei Dou Shu, Onomancy (Name Study), and MBTI using n8n workflows and Python algorithms. We will complete a production-level service system from start to finish where 13 specialized agents collaborate to generate professional-grade Saju reports.

This is a tutorial-style Claude Code development course that beginners and those new to development can easily follow. Based on experience teaching planners, designers, and non-developers in offline classes, the course covers everything in one continuous flow—from beginner examples not covered in books to intermediate development examples. Using Claude Code, you will experience the step-by-step development process, starting with Hello World and progressing to a lucky draw, a device-optimized café menu management web page integrating Claude Design with Excel data, and a SaaS web application that can be connected to a database and deployed to cloud services. Tutorials are provided at each stage—theory, practice, and projects. Through hands-on exercises across 6 levels, you will learn AI-collaborative development methodologies and experience the entire development cycle, including cloud deployment and operations.

We will develop a web app that collects data using concurrent/parallel programming in Python, loads it into MongoDB, analyzes it, and visualizes the collected data using FastAPI, a modern web framework. In the process, you will learn and understand Python concurrent/parallel programming using coroutines, multi-threading, and multi-processing, and understand how to use them to improve Python's performance. We will help you apply Python concurrent/parallel codes directly to your work. We will not only teach you how to develop and end, but also how to deploy and operate an actual app on the AWS cloud.

Are you overwhelmed by AI-generated backend code and unsure where to look? Learn how to identify missing conditions in requests, APIs, data, concurrency, authentication, and caching—and what requirements to give the AI and what criteria to use when reviewing its output.

From analyzing the encryption mechanisms of ZIP, PDF, and MS Office files to hands-on dissection of the WannaCry ransomware, master static analysis tools and GPU-accelerated cracking techniques. This is a practical, security engineering-focused course that builds incident response capabilities through the development of next-generation malware analysis pipelines using OpenAI Codex and automated report generation.

I sent the same request twice and got different answers. Which one can I use in my program? We’ll walk through actual API call logs on screen, step by step, examining the structure of requests and responses, structured outputs, streaming, tool calls, errors and retries, and costs. Instead of treating the mere arrival of text as success, we’ll establish criteria for accepting or holding a response based on its status, format, evidence, and cost.