Slide 16 of 27
Part 3 · ScenariosSlide 16
Slide 16 · OWASP Scenario #3
Data Leak via Training Data.
📄 OWASP LLM Top 10:2025 · LLM02 Scenario #3
Scenario 3 of 3
“Negligent inclusion of sensitive data in training sets results in disclosure through the model's outputs.”
This is the long-tail version of Slide 13's poem attack — and also a forward-looking risk from Slide 12's Samsung incident. Any sensitive text an organization feeds into a model's training or fine-tuning pipeline becomes part of what that model might say back, to someone else, later.
The uncomfortable part: by the time this scenario surfaces, the sensitive data has often already been baked into model weights. Unlike a database record, you can't simply delete it — which is exactly why OWASP's mitigations for this scenario focus on prevention (sanitize before training) rather than cleanup after.
← BackNext → The Pattern