HighMemorization●Realized
PII memorization & regurgitation
Privacy & Data ProtectionDescription
The model memorises and reproduces verbatim personal or confidential data from its training/fine-tuning set.
Example scenario
A fine-tuned model trained on support tickets reveals a real customer's account number in an answer.
Real-world evidence●Realized
In March 2023 an OpenAI production bug exposed ChatGPT users' conversation titles and fragments of other users' chat histories in real time. Separately, Samsung employees inadvertently submitted proprietary source code and meeting notes to ChatGPT, causing confirmed data exfiltration through an AI system in production.
Tested by FinProof
Data rights
Benchmarked by FinProof Primary mitigations
- PII scrubbing pre-training
- differential privacy
- output PII filters
- memorization testing
- deduplication.
Detection signals
Canary/extraction testing; PII-in-output scanning; memorization benchmarks.
Mitigating controls
6 Non-agentic controls
ZYC-PRIV-001Non-Agentic
PII Detection and Classification
Privacy & Data Governance
ZYC-SEC-001Non-Agentic
Adversarial Testing
Security
ZYC-LEGAL-001Non-Agentic
AI Regulation Compliance
Legal Compliance
ZYC-PRIV-011Non-Agentic
Differential Privacy
Privacy & Data Governance
ZYC-PRIV-012Non-Agentic
Synthetic Data
Privacy & Data Governance
ZYC-PRIV-008Non-Agentic
PII Field Identification Accuracy
Privacy & Data Governance