Observed incident

Cloudflare GLM-5.3 responses exposed other customers’ request content

Reported

Published here

What happened

Cloudflare confirmed 155 affected responses during August 27–September 9. Overlong GLM-5.3 prompts reached stale cache data from other requests; the provider found no evidence of deliberate exploitation.

Parties and reporting

AI tools and systems: Workers AI @cf/zai-org/glm-5.3, SGLang disaggregated prefill/decode, Mooncake KV transfer

Outcome

Cloudflare corrected capacity and rejection handling and restored the model September 9; no other Workers AI model was affected. October 26 reports recommend SGLang 0.5.22 or later for broader hardening.

Operating lesson

Test cross-tenant isolation during oversized requests and failures, and verify cache capacity after every model-serving change.

Evidence and limits

155 is a response count, not a customer count. This is confirmed accidental exposure, not established malicious theft. October 26 disclosed customer impact; the same engine bug had a July 6 public issue, while September 9 status updates described availability errors.

Sources