This course introduces developers, engineers, and security practitioners to the OWASP LLM Top 10 and secure coding practices for LLMs. Learning Objectives: Understand the top security risks specific to LLMs. Recognise real-world attack scenarios (e.g., prompt injection,sensitive data disclosure), Learn defensive strategies and coding practices to mitigate risks, and Gain hands-on experience via the SecDim platform.
This module introduces developers, engineers, and security practitioners to the OWASP LLM Top 10 and secure coding practices for Large Language Models (LLMs).
Prompt injection is one of the most critical risks identified in the *OWASP LLM Top 10*.
Complete this lab to learn how prompt injection works in practice
Sensitive information disclosure occurs when an LLM unintentionally exposes confidential data during interaction.
This bot is leaking secrets, let's fix it.
Unlike traditional software, LLM development depends heavily on *third-party models and datasets*
After loading the model, a command was executed on the machine. Let's find and fix the issue.
Data poisoning is when an attacker sneaks bad or biased examples into the data used to train or fine-tune a model.
“Improper Output Handling” refers to situations where the output from an LLM is passed downstream into web pages, system calls, code execution, SQL queries, file paths, or other components, without sufficient validation, sanitization, or context-aware handling.
The output from this chatbot is executing some code. Let's find out how.
Excessive agency happens when the model has more powers than it needs — violating the principle of least privilege.
This bot can do too much and in the wrong way.
System Prompt Leakage is the situation in which the “system prompt” (the hidden instructions given to a large-language-model application) is exposed either via direct extraction or indirect inference.
Vector and embedding systems are a cornerstone of many modern LLM applications: they power semantic search, retrieval-augmented generation (RAG), similarity matching, and much more.
The models we build to assist, summarise or generate content can themselves become vectors for false, misleading or fabricated information—often with credible language and confident tone.
Unbounded Consumption is scenarios in which users or attackers consume model resources in uncontrolled ways: making unlimited inference requests, generating endless output, or triggering costly compute loops.
This chatbot is burning our bill and CPU. Let's do something.
To obtain the certificate, complete the following challenge.
