Cross-request KV cache exposure in the GLM-5.3 Workers AI model, and our remediation with SGLang
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Cloudflare disclosed that between August 27 and September 9, 2026, 155 responses from its GLM-5.3 Workers AI model leaked data from other customers' requests, affecting only prompts over 681,280 tokens.
Why it ranked
No. 24 of 124 with a score of 0.67. The panel rated it highest on technical depth (0.96) and substantiated (0.92), lowest on audience breadth (0.58). The claim check took 0.100 off.
Every answer behind this is in the table under Score.
Score
0.67total under rubric v2.2
- Significance0.68
- Reader interest0.88
- Audience breadth0.58
- Practical relevance0.88
- Substantiated0.92
- Technical depth0.96
- New information0.92
- Counts against: Promotional0.09
- Counts against: Overclaim0.13
Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).
| Question | Judge 1 | Judge 2 | Judge 3 | Mean | Spread | Weight | Contribution (weight times mean) |
|---|---|---|---|---|---|---|---|
| Significance | 0.66 | 0.55 | 0.82 | 0.68 | 0.27 | +0.24 | +0.163 |
| Reader interest | 0.84 | 0.85 | 0.95 | 0.88 | 0.11 | +0.26 | +0.229 |
| Audience breadth | 0.44 | 0.60 | 0.70 | 0.58 | 0.26 | +0.13 | +0.075 |
| Practical relevance | 0.94 | 0.80 | 0.90 | 0.88 | 0.14 | +0.22 | +0.194 |
| Substantiated | 0.94 | 0.85 | 0.98 | 0.92 | 0.13 | +0.05 | +0.046 |
| Technical depth | 1.00 | 0.90 | 0.98 | 0.96 | 0.10 | +0.05 | +0.048 |
| New information | 0.90 | 0.90 | 0.97 | 0.92 | 0.07 | +0.05 | +0.046 |
| Promotional | 0.05 | 0.20 | 0.03 | 0.09 | 0.17 | −0.19 | −0.018 |
| Overclaim | 0.15 | 0.20 | 0.04 | 0.13 | 0.16 | −0.11 | −0.014 |
| Sum of contributions | +0.769 | ||||||
| Claim-check penalty | −0.100 | ||||||
| Total, as published | 0.669 | ||||||
Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.
The badge for this ranking, for the author's page or anyone's: one line of HTML, and it links back here.
<a href="https://rubric.hrih.si/editions/2026-10-27/24/"><img src="https://rubric.hrih.si/badge/2026-10-27/24.svg" alt="Ranked #24 on Rubric, 27 Oct 2026" width="245" height="28"></a>Claim check
Claim check: 2 matches source · 1 not in source · 4 no source given · 1 couldn’t check. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
No source givenNumber
Between August 27 and September 9, 2026, 155 GLM-5.3 Workers AI responses were affected, all to prompts longer than 681,280 tokens.
The article gives no link or citation for it.
The article says “In total, 155 responses were affected, all to prompts longer than 681,280 tokens”
The source says No source to compare with.
No source givenNumber
Cloudflare deployed GLM 5.3 with a 1,048,576-token context window, but the KV cache pool on its prefill GPUs only had room for 681,280 tokens.
The article gives no link or citation for it.
The article says “We deployed GLM 5.3 with a 1,048,576-token context window, but the KV cache pool on our prefill GPUs only had room for 681,280 tokens”
The source says No source to compare with.
Not in sourceNumber
Cloudflare took the GLM 5.3 model offline within 40 minutes of declaring an incident.
The cited source does not contain it.
The article says “took the model offline within 40 minutes of declaring an incident”
The source says “Cloudflare is aware of, and investigating, increased Workers AI errors when attempting to use GLM 5.3. Updates to follow”
Source cited by the article: cloudflarestatus.com/incidents/jsvvrtk5hvym
Matches sourceDate
The SGLang rejection bug was reported to SGLang by another user as issue #30233 on July 6, 2026.
The cited source says the same thing.
The article says “This bug had been reported to SGLang by another user as issue #30233 on July 6, 2026”
The source says “closed, opened 2026-07-06 by kevin-14”
Source cited by the article: github.com/sgl-project/sglang/issues/30233
No source givenVersion
The SGLang patch for this issue was released in v0.5.20, with follow-ups in v0.5.21 and v0.5.22.
The article gives no link or citation for it.
The article says “The patch for this issue was released in SGLang v0.5.20, with follow-ups in v0.5.21 and v0.5.22”
The source says No source to compare with.
Couldn’t checkDate
The fix in SGLang PR #36988 was merged September 6 for a related failure path but not yet released at the time of the incident.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “The fix was already on SGLang main in #36988, merged September 6 for a related failure path but not yet released”
The source says No passage could be compared.
Source cited by the article: github.com/sgl-project/sglang/pull/36988
No source givenFirst, only or fastest
GLM 5.3 was the only model on Workers AI affected by this bug.
The article gives no link or citation for it.
The article says “GLM 5.3 was the only model affected by this bug”
The source says No source to compare with.
Matches sourceNumber
Cloudflare's page-tag integrity prototype cost under 1% in throughput and tail latency in its tests.
The cited source says the same thing.
The article says “In our tests it cost under 1% in throughput and tail latency”
The source says “The cost is under 1% on both throughput and tail latency, and even the upper bound of the 95% confidence interval stays near 1%”
Source cited by the article: blog.cloudflare.com/smaller-faster-safer-models
Also covered by
Elsewhere
Elsewhere: HN 111 pts, front page
Recorded for context only, checked 27 Oct 2026, 05:46 UTC. It played no part in the ranking.