THE HOOK
Cloudflare's hosted AI briefly played pass-the-parcel with other customers' data.
THE GIST
Cloudflare disclosed that its GLM-5.3 Workers AI service sent some long requests cached information from other customers, per The Neuron's roundup. A configuration mismatch triggered an SGLang bug, hitting 155 responses between August 27 and September 9. Both the config and the inference engine are now patched.
THE TAKE
Credit for disclosing it. But this is the nightmare everyone waved away when they piled onto shared inference: caching is how you make tokens cheap, and caching across tenants is how you leak. 155 responses sounds small. It isn't small if one of them was yours.
OPINION
Cloudflare says it found no sign of deliberate exploitation, and the two-week window plus open disclosure is better than most vendors manage, as The Neuron notes. The bigger lesson lands on every host running SGLang or similar engines with prefix caching: tenant isolation has to be tested, not assumed.
- Story type: Breaking news
- Impact: Medium
- Confidence: Medium (multiple reports)
- Audience: Developers
- What changed: Cloudflare disclosed 155 Workers AI responses leaked cached data across customers via an SGLang bug, now patched.
- What to watch: Whether other SGLang-based inference hosts audit their prefix caches and disclose similar issues.
- Reading time: 1 minute
