CYBRARY Who Was in the Training Set?
A regulator asks whether a customer's records were used to train the production model, and whether the model leaks them. Run the membership-inference and memorisation tests, then decide what changes.
DSAR-1102 — 'confirm whether my data trained your model'
Privacy engineer
ML engineer
Legal
45 minutes · advanced
Execution trace
00:00:00enterprisetraining-pipeline
attack step
Corpus assembled from production tables
Includes free-text support notes containing PII.
00:01:00modelassistant@2026.04
attack step
Canary string recalled verbatim
A unique planted record is reproduced on prompting.
canary=hit · exact_match=true
00:02:00modelmembership-probe
attack step
Members separable from non-members
Loss-based probe achieves 0.71 AUC — well above chance.
00:03:00enterprisedata-catalog
attack step
No lineage recorded
Nobody can answer the regulator's question from records.