Sensitive information leakage through model outputs

https://taxonomy.eticas.ai/risk/sensitive-information-leakage

Maturity: established

Personal or confidential/proprietary information present in training data, inference-time context, or connected systems, surfacing in model outputs through memorisation, context disclosure, cross-tenant contamination, or system-prompt extraction. Broadens the former pii-leakage (personal data only) to also cover organisational and commercial confidentiality (the former confidential-information-leakage, retired and absorbed here).

Also known as: PII leakage rate · Privacy leakage

System type: ADM and LLM systems
Lifecycle stages: In Processing, Post Processing

Mappings to external frameworks

Standards & frameworks

Framework Reference
EU AI Act (Regulation 2024/1689) Article 15(5) — cybersecurity (confidentiality attacks)
AIUC-1 — AI Underwriting Company Standard Prevent PII leakage
NIST AI 600-1 — Generative AI Risk Profile Data Privacy (leakage clause)

Taxonomies & vocabularies

Framework Reference
MIT AI Risk Repository Compromise of privacy by leaking or inferring sensitive information
W3C Data Privacy Vocabulary — AI Extension Unauthorised Data Disclosure
AIR 2024 Privacy → Unauthorized Privacy Violations × PII
IBM AI Risk Atlas Output → Revealing personal information in output

Source: HRA project taxonomy