Definition

Security property in shared inference and SaaS platforms ensuring one customer’s data, prompts, and model state cannot influence another customer’s requests or responses.

Key Points

  • 2026-10-26: cloudflare disclosed cross-tenant kv-cache leakage on glm-5-3 Workers AI when prompts exceeded misconfigured pool size—155 responses Aug 27–Sep 9, 2026 (2026-10-26-cloudflare-glm-kv-cache-exposure)
  • Failure mode: disaggregated prefill/decode plus rejected over-length requests continuing with placeholder tokens left stale cache pages from prior tenants (sglang)

Sources