Training Data Replay CLOZE [MASK] Elicitation - ATR-2026-00401 (aeb462a8-4530-5539-9525-41bdc6506c7e)
Detects requests to fill in [MASK] tokens from training data verbatim, from NVIDIA garak leakreplay probe. The LiteratureCloze/PotterCloze/NYTCloze/GuardianCloze probes send a prompt template asking the model to recall a proper name from its training data using [MASK] fill-in-the-blank format, eliciting verbatim copyrighted text reproduction. Key signals: [MASK] combined with "training data" recall framing, "exactly one word long" constraint,