NEWPractise supported commands in browser simulations, alongside the lesson.Try a browser practice exercise.Try a lab♥ Codelivly is 100% free, no paywall. If it helped you, chip in to keep it running.100% free. Help keep it so.Support us
Networking Fundamentals
You cannot secure a network if you do not understand how data moves across it. The Networking Fundamentals learning path
ViewEndpoint & Network Security Monitoring
Dive into the critical telemetry that powers modern threat detection. The Endpoint & Network Security Monitoring module
ViewSOC Tools: Hands-On Practicals
Theory is useless without muscle memory. You can read a hundred books on how a SIEM works, but until you’ve stared down
ViewExploitation Fundamentals
Amateurs stop at the proof of concept. They pop an alert(1), dump a database via SQLi, or find an IDOR, write a report,
ViewRouting & Switching Fundamentals
Build a rock-solid foundation in modern network infrastructure. This learning path takes you deep into the mechanics of
ViewNetwork Design and Architecture
Elevate your career from network technician to strategic network architect. This advanced learning path takes you beyond
ViewPreview · 30 min
Learn how LLMs leak training data like PII and secrets. Master defenses against data reconstruction attacks with pre-training sanitization and Output DLP.
In the previous lesson, RAG leaked data by processing confidential documents at run time. What happens when the sensitive data is not in an external document but permanently baked into the model weights?
Training Data Leakage is a severe subset of Sensitive Information Disclosure (OWASP LLM06). An LLM memorises and repeats sensitive information, such as PII, hard-coded secrets or proprietary source code, that was in its pre-training or fine-tuning data.
Unlike a database breach, it does not use a flaw in a query. It uses the statistical nature of the model. If a unique string (an internal API key or one employee email) appeared often or uniquely enough in training, the model may "overfit" and memorise it.
For a defender this is a hard case: you cannot patch the weights of a trained model. Once a model is deployed with contaminated data, the information is embedded. Defence needs strict data governance before training and output filtering at run time.
All scenarios, code and targets in this lesson are fictional and made for isolated training labs.
Create a free account to unlock the rest of this task, complete tasks, and earn Rep.
Sign in