“Implement data sanitization to prevent user data from entering the training model. This includes scrubbing or masking sensitive content before it is used in training.”

“Apply strict input validation methods to detect and filter out potentially harmful or sensitive data inputs, ensuring they do not compromise the model.”

The Samsung leak (Slide 12) is exactly the gap this closes — there was no sanitization step between an engineer's clipboard and ChatGPT's input field. No mask, no filter, no warning. The poem attack (Slide 13) shows the training-side half: whatever wasn't scrubbed before training became extractable later.

→ Mask or scrub PII, credentials, and business identifiers before any text reaches a training or fine-tuning pipeline
→ Run input validation on what users paste in, not just what they type — copy-pasted code and documents carry the same risk
→ Treat “scrubbing before training” and “filtering at input time” as two separate, both-required steps

Try pasting synthetic PII or fake credentials into your own app's input field. If nothing flags it, scrubs it, or warns the user, this control doesn't exist yet.