LowAttribute Inference●Realized
Sensitive Attribute Deduction
Privacy & Data RiskDescription
Adversary infers protected attributes (race, health status, political views) from model outputs or query patterns. Enables discrimination at scale without direct data access.
Example scenario
Insurance company queries agent to infer health risk attributes from behavioural data, enabling discriminatory pricing without direct health data.
Real-world evidence●Realized
Multiple data protection authorities in the EU have issued enforcement actions against AI systems processing personal data without a valid legal basis or adequate consent mechanisms, with Italy's ChatGPT ban and subsequent CNIL and Irish DPC investigations representing confirmed regulatory findings of this failure in production.
Primary mitigations
- Attribute inference testing
- fairness-aware training
- output correlation monitoring.
Detection signals
Attribute inference accuracy; model output fairness metrics.
Mitigating controls
6 Dual coverage
ZYC-PDP-03Agentic
Memorisation Testing & Output PII Scrubbing
Privacy & Data Protection
ZYC-PDP-05Agentic
Attribute-Inference Mitigation
Privacy & Data Protection
ZYC-PRIV-001Non-Agentic
PII Detection and Classification
Privacy & Data Governance
ZYC-PRIV-002Non-Agentic
Sensitive data leakage
Privacy & Data Governance
ZYC-PRIV-005Non-Agentic
Consent Scrub
Privacy & Data Governance
ZYC-PRIV-006Non-Agentic
Data Minimization
Privacy & Data Governance