Cloudflare discloses that GLM-5.3 on Workers AI leaked other customers’ text in 155 responses via stale KV cache (Aug 27–Sept 9)
Confirmed
The takeaway
Cloudflare published on Oct 26, 2026 a post-mortem showing that very long prompts (over 681,280 tokens) to the GLM-5.3 model on its Workers AI platform could return text derived from other customers’ requests between Aug 27 and Sept 9, 2026.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 2 of 5
- Last verified
- 27 October 2026
Your AI and this story
- GPT-6.2 Astra148 days after its cutoff
- Claude Opus 5.5118 days after its cutoff
- Gemini 4 Hydrogen209 days after its cutoff
- Grok 4.7148 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 118 days before it.
Key facts
- Model: GLM-5.3 (@cf/zai-org/glm-5.3) on Cloudflare Workers AI; Cloudflare says no other Workers AI model was affected
- Exposure window: Aug 27 to Sept 9, 2026; 155 responses, all to prompts longer than 681,280 tokens; no evidence of deliberate exploitation (Cloudflare)
- Cause: the prefill KV pool held 681,280 tokens while the model’s context window was 1,048,576; SGLang replaced a prompt that did not fit with a one-token placeholder instead of returning an error, and the decode node generated from stale KV pages left by earlier requests
- Timeline: HackerOne report Sept 8, 23:11 UTC; model made private Sept 9, 14:00 UTC; restored 19:40 UTC; post published Oct 26
- Fix: KV pool raised to about 2.04M tokens; over-length requests now rejected before KV transfer (in SGLang v0.5.20+, v0.5.22+ recommended); deployment checks of pool size vs context length; work on page-level integrity tags and formal verification of KV transfer
- Reach: 145 points on the Hacker News front page (Oct 26)
What happened
Disaggregated inference splits prompt processing (prefill) and generation (decode) across machines that hand over KV-cache pages. When a 1M-token model was deployed with a smaller prefill pool, the serving engine failed silently on the longest prompts, and the decoder produced output from whatever cache pages earlier requests had left behind, which could include other users’ data.
Why it matters
Million-token contexts are now routine, and cross-tenant leakage through shared inference caches is a new class of incident for hosted open-weight models. The fix went upstream into SGLang, which many providers use.
Sources
2 sources from 2 sites. Numbers match the chips in the text.
2 sources: 1 primary, 1 reaction
Primary
- Cloudflare blog: cross-request KV cache exposure in the GLM-5.3 Workers AI modelblog.cloudflare.com, official
Reactions
- Hacker News discussionnews.ycombinator.com, discussion
Changes
- Filed (Cloudflare post read directly)