Cloudflare and SGLang disclose cross-request cache exposure

Cloudflare and the SGLang team disclosed a fixed Workers AI isolation failure that put content from other requests into 155 GLM-5.3 responses. The exposure ran from August 27 to September 9; Cloudflare reported no evidence of deliberate exploitation.

What the public record establishes

The October 26 reports are a coordinated public disclosure. Cloudflare dates the production exposure to August 27 to September 9 and says it restored the model on September 9 after correcting configuration and request-rejection handling. The count of 155 affected responses is not a count of customers. [1][2]

Cloudflare’s prefill cache capacity was smaller than the configured context window. A rejected request could still enter SGLang’s disaggregated handoff and read cache data belonging to another request. The SGLang team documented joint remediation with Cloudflare; neither report identifies other Workers AI models as affected. [1][2]

Sources

  1. Cross-request KV cache exposure in the GLM-5.3 Workers AI model, and our remediation with SGLang

    Cloudflare · Oct 26, 2026

  2. SGLang security update: improving request isolation

    LMSYS Org · Oct 26, 2026

  3. When the Inference Engine Leaks: Cross-Tenant KV Cache Exposure

    Cloud Security Alliance · Nov 2, 2026