• views: 40
• date: 20 August 2026
• posted by: voska89

What you'll learn

⚡ Identify OWASP LLM04 poisoning risks across training data, fine-tuning datasets, model artifacts, and RAG ingestion pipelines.
⚡ Design repeatable QA tests using clean baselines, poisoned inputs, trigger conditions, and multiple test runs.
⚡ Detect model backdoors by comparing normal and triggered behavior, even when standard benchmark results appear unchanged.
⚡ Test RAG ingestion gates to verify altered or untrusted content is quarantined before it reaches the vector store.
⚡ Validate dataset, model, and RAG provenance using hashes, signatures, metadata, source controls, and behavioral evidence.
⚡ Document poisoning findings with reproducible evidence, containment actions, and release or ingestion gate recommendations.

Requirements

❗ No prior AI security or machine learning security experience is required.
❗ Basic familiarity with software testing or QA concepts is helpful, but beginners can follow the course.
❗ General familiarity with LLM applications is useful. RAG, model poisoning, and provenance concepts are explained in the course.
❗ You need a computer with internet access and a modern web browser. No paid AI tools or specialized hardware are required.

Description

Data and model poisoning can compromise an AI system long before a user submits a prompt. A poisoned training record, altered model artifact, or untrusted RAG document may introduce behavior that ordinary functional tests and benchmark scores fail to expose.
This course shows QA engineers, software testers, SDETs, developers, and application security professionals how to test these risks systematically. You will learn where poisoning can enter the AI lifecycle and how to build evidence that distinguishes expected model variation from a repeatable security failure.
The course covers poisoning risks in training and fine-tuning data, pretrained model artifacts, and retrieval-augmented generation pipelines. You will learn how to establish a clean baseline, introduce controlled test conditions, repeat probes, compare normal and triggered behavior, and record results that another tester can reproduce.
Two practical demonstrations connect the testing methods to realistic systems. The first examines a poisoned model that behaves normally under standard prompts but changes behavior when a hidden trigger appears. The second follows an altered document through a RAG ingestion pipeline and shows how provenance checks and ingestion gates can quarantine it before it reaches the vector store.
You will also learn how hashes, signatures, metadata, source controls, benchmark results, and behavioral testing fit together. No single signal proves that a dataset or model is trustworthy. The goal is to combine those signals into defensible release and ingestion decisions.
By the end of the course, you will be able to design a focused OWASP LLM04 test plan, collect reproducible evidence, document poisoning findings, and recommend appropriate containment and gate controls. No prior AI security or machine learning security experience is required.

Who this course is for

⭐ QA engineers, software testers, and SDETs who want practical techniques for testing LLM applications for data and model poisoning.
⭐ Test leads and QA managers responsible for AI security test plans, release criteria, and reproducible evidence.
⭐ Developers and ML engineers working with fine-tuning datasets, model artifacts, or RAG ingestion pipelines.
⭐ Application security professionals who want QA-focused coverage of OWASP LLM04 risks and controls.
⭐ Software testing professionals moving into AI security who do not yet have specialized machine learning security experience.

Homepage

https://www.udemy.com/course/ai-security-testing-llm-04-data-and-model-poisoning

   
   
   

Forex Trading

Crypto Trading. 20 USD Bonus after KYC

We need your support!
Make a donation to help us stay online
        
Bitcoin (BTC)
bc1q08g9d22cxkawsjlf8etuek2pc9n2a3hs4cdrld
	
Bitcoin Cash (BCH)
qqvwexzhvgauxq2apgc4j0ewvcak6hh6lsnzmvtkem

Ethereum (ETH)
0xb55513D2c91A6e3c497621644ec99e206CDaf239

Litecoin (LTC)
ltc1qt6g2trfv9tjs4qj68sqc4uf0ukvc9jpnsyt59u

USDT (ERC20)
0xb55513D2c91A6e3c497621644ec99e206CDaf239

USDT (TRC20)
TYdPNrz7v1P9riWBWZ317oBgJueheGjATm

0 1 2 3 4 5

| Print version |

	Information
		Users of GUESTS are not allowed to comment this publication.