Skip to content
REDMAW

AI Systems

Training-Data Extraction

Adversarial interaction attempts to make a model reveal memorized or sensitive information associated with its training or fine-tuning context.

RedMaw

Definition

Models can retain specific content from their training or fine-tuning data. Extraction is the attempt to elicit that retained material through adversarial interaction.

Why fine-tuning raises the stakes

Organizations fine-tune on their own data, and that data is frequently operational: support conversations, internal documentation, customer records. Where such material is memorized, the model becomes a disclosure surface for it, reachable by anyone who can send it a prompt.

What reduces exposure

  • Treating training-data selection as a data-protection decision
  • Removing or minimizing sensitive records before fine-tuning
  • Testing for extraction after each fine-tune, against known-sensitive content
  • Retrieval with access control, in place of embedding data into weights
Where it leads
  1. Adversarial Prompt
  2. Model
  3. Memorized Content
  4. Disclosure

Each hop validated by successful exploitation

Tags

  • disclosure
  • privacy
  • fine-tuning

Validation

How RedMaw validates this attack

Validation runs only inside authorized scope, with agreed exploitation limits.

  1. 01Establish what sensitive content the training or fine-tuning context may contain
  2. 02Attempt elicitation of that content through adversarial interaction
  3. 03Determine whether disclosure is reproducible
  4. 04Re-test after each fine-tune against the accepted baseline

Keep going

Related

Capabilities

Use cases

Stop assuming you are secure. Prove it.

Continuously test what an attacker can actually reach across your applications, SaaS identities, internal infrastructure and AI systems.

Request a demoExplore the platform