What happened
On Aug 27 2026 Cloudflare deployed the GLM‑5.3 model with a 1 048 576‑token context window while the underlying SGLang inference engine’s KV‑cache could only hold 681 280 tokens. When prompts exceeded the cache, SGLang incorrectly transferred the KV cache from a prior request to the next request instead of aborting. Between Aug 27 and Sep 9, 155 long‑prompt requests (≥ 681 280 tokens) received responses containing fragments of other customers’ data. Cloudflare was notified on Sep 8 by a researcher, took the model offline within 40 minutes, applied a corrected configuration and a patched SGLang version, and restored service on Sep 9. No evidence of deliberate exploitation was found, and no other Workers AI models were affected.
Why it matters for trust and compliance
- The incident demonstrates that without continuous verification of tenant‑isolation settings, even a short‑lived misconfiguration can breach data‑segregation controls required by multiple frameworks.
- Shows the need for continuous monitoring of configuration drift that could erode isolation guarantees.
- Provides audit‑ready evidence that segregation controls are enforced across all AI inference workloads.
Who is affected
TECH_SAAS CLOUD_HOST
Recommended actions
- Audit KV‑cache and context‑window settings for all multi‑tenant AI models.
- Implement automated alerts for requests that exceed resource limits.
- Collect and retain logs proving tenant isolation for audit readiness.