Threat
Cross-request KV cache exposure in the GLM-5.3 Workers AI model, and our remediation with SGLang
First reported cloudflare.com
Page published
Earliest dated coverage: 27 Oct 2026 · First observed: 27 Oct 2026 · Latest dated coverage: 27 Oct 2026
Coverage timeline
Single-source incident — one report is available.
Why it matters
The Workers AI incident demonstrates that inference-engine error handling and stale KV-cache state can expose customer data across tenant boundaries even without deliberate exploitation.
Cloudflare confirmed that a tenant-isolation flaw in Workers AI's GLM-5.3 model exposed other customers' request data in 155 responses between August 27 and September 9, 2026. Prompts exceeding 681,280 tokens triggered an SGLang rejection-handling bug that allowed generation to use stale KV-cache pages. Cloudflare corrected the configuration, patched the inference engine, and restored service, reporting no evidence of deliberate exploitation.