1min.

1MIN · AI & AGENTS

A rejected model request crossed an isolation boundary in Workers AI

Cloudflare reports that a rejected long prompt still reached the decoding stage of its GLM-5.3 service, which read cache pages the request had never written. Those pages contained earlier users’ data. The company identified 155 affected responses during August 27–September 9 and says it remediated the issue on September 9; its October report details the failed handoff and subsequent hardening.

Cloudflare Security · Oct 26, 2026

Why it matters

Failure and cancellation paths need the same ownership checks as successful transfers when memory is reused between tenants.

More context

The prefill stage rejected work but decode accepted its handoff and reserved more pages than were initialized. Cloudflare added admission fixes and stronger page-lifetime and transfer checks, and recommended SGLang 0.5.22 or later in the report. This is a catch-up engineering postmortem, not a new ongoing incident.

Keep in mind

The scope and absence of detected deliberate exploitation are Cloudflare’s findings. Other deployments need their own version, configuration and isolation review rather than assuming the same exposure.

Original read

More engineering ideas, one minute at a time.

Explore more stories