Embedding / vector-store leakage (cross-tenant)
Data & Input IntegrityDescription
Weak isolation or invertible embeddings in a shared vector store leak one tenant's or user's sensitive content to another.
A multi-tenant advisory platform returns chunks of another bank's confidential memo via the shared index.
While general IT cascading failures (e.g., CrowdStrike 2024, power grid interdependencies) are well-documented, a confirmed production incident where an AI system's failure specifically cascaded through interconnected critical infrastructure causing widespread disruption has not been publicly attributed. The risk is plausible given AI's growing role in grid management and SCADA systems but remains prospective.
No public incident on record — evidence level: Theoretical
Primary mitigations
- Per-tenant namespace isolation
- access control on vector DB
- embedding inversion testing
- encryption at rest
- metadata filtering.
Detection signals
Cross-tenant retrieval tests; access-log anomaly detection; embedding-inversion red-team.