Advanta is currently undergoing final system calibration ahead of launch. Selected infrastructure and experiences may still be in active refinement.

advanta

RC-2 · Risk Taxonomy 2026.1

Data leakage

Mechanism

Information that should have stayed inside the function reaches a model provider, vendor infrastructure, or another customer via AI data pathways. Three variants: deliberate-by-design (vendor terms permit prompt-to-training), inadvertent (privileged content captured in vendor logging), operational (vendor support staff access prompt history without per-incident customer authorisation).

Evidence (what the Evidence Register holds)

Vendor DPA confirming data isolation from training; residency confirmation; vendor employee access control documentation; prompt-log retention bounds; per-incident vendor access logs.

Mitigation

Tenant-bound deployments for sensitive matters; prompt-data minimisation; vendor DPA enforcement (Chapter 6); residency configuration; vendor employee access controls.

Editorial Framing

Data leakage is Class 2 because it is the upstream mechanism for several downstream classes — most notably Client confidentiality breach (RC-7) and Regulatory non-compliance (RC-5). A clean vendor DPA + tenant-bound deployment + residency configuration is the structural mitigation.

Indicative Examples

  • Vendor permitted prompt-to-training under default terms
  • Privileged content captured in vendor logging
  • Vendor support staff accessing prompt history without authorisation