ZDR is per-layer: what each provider's privacy policy actually says in writing
Mission-Leave-6551 · reddit · 2026-10-29
'No training, no storage, no human review' are three separate promises, and most layers only make one or two. The post breaks down each layer: DeepSeek's Feb 2026 policy lists training on your inputs with opt-out on request and China storage, while z.ai claims no content storage and Singapore processing without addressing training. Host docs' exceptions matter (metadata logging, short image retention, encrypted bulk data). Router ZDR toggles only filter upstreams—the router still sees prompts—and one major router dropped Google models from ZDR routing on Oct 18. Harness/telemetry risks remain: zcode reportedly uploaded entire workspaces.
More from Safety
- XBOW's AI agent finds critical unauthenticated RCE in EAC voting calculator, fixed before midterms — moyix · 2026-10-29
- Openly available GLM model completes multi-stage cyberattacks in controlled tests — S_OhEigeartaigh · 2026-10-29
- Security researcher questions why frontier AI labs are investing heavily in vulnerability discovery — matthew_d_green · 2026-10-29
- Banks grapple with detecting third-party AI agents using legitimate customer sessions — CommonGrapefruit3653 · 2026-10-29
- Anti-bot scripts hide ricin-dose function names to trip AI safety filters — matthew_d_green · 2026-10-29
- OCR plus PII model redacts patient data burned into X-ray pixels on one GPU — MaziyarPanahi · 2026-10-29