Skip to the content
AI and Cybersecurity Learning and Practice Center Test automation Open the Learning Center

WeaknessASC-2026-0030Other weakness

Memorized training data and secrets in model output

MediumWorkarounds only

Models can repeat fragments of their training data, including personal data or secrets that were in it, when prompted the right way.

How to detect it

  • Outputs that contain real names, keys or verbatim passages
  • Fine-tuning data that was not scrubbed of secrets

Workarounds

  • Scrub secrets and personal data from fine-tuning sets
  • Filter outputs for secret patterns and known sensitive strings

Memorization cannot be removed from a trained model without retraining.

How to investigate

  1. Find the prompt that produced the leak and test variants
  2. Rotate any secret that appeared and review the training data source

Sources

Open it in the Learning Center's vulnerability database