Cloudflare discloses GLM-5.3 cross-tenant leak and fixes it

1. 22:19
   X@Pick_Analytica· Discussion

   Source: https://blog.cloudflare.com/kv-cache-exposure-remediation/(opens the original post in a new tab)
2. 22:17
   X@Pick_Analytica· Discussion

   Cloudflare says its Workers AI GLM-5.3 model sent 155 responses with other customers' data from Aug 27 to Sept 9, all o…(opens the original post in a new tab)
3. 22:15
   X@heypromptpower· Discussion

   For 13 days, a long enough prompt to GLM-5.3 on Cloudflare Workers AI could return another customer's data: names, invo…(opens the original post in a new tab)
4. 20:38
   X@AiToolsObserver· Discussion

   Cloudflare’s GLM-5.3 Workers AI leak shows how KV cache and disaggregated SGLang configs can quietly expose prompts. He…(opens the original post in a new tab)
5. 16:30
   X@thesystemprompt· Discussion

   Cloudflare discloses a cross-request KV cache leak: 155 GLM 5.3 responses on Workers AI (Aug 27-Sep 9) contained other…(opens the original post in a new tab)
6. 15:33
   X@AndroOxinu· Discussion

   Cross-request KV cache exposure in the GLM-5.3 Workers AI model, and our remediation with SGLang https://share.google/Y…(opens the original post in a new tab)
7. 13:21
   X@pramod_kodag· Discussion

   cloudflare disclosed a bug in workers AI's GLM-5.3 service where some long requests got cached data from other customer…(opens the original post in a new tab)
8. 11:30
   X@AbhishekShende_· Discussion

   Cloudflare post-mortem: an LLM on Workers AI returned other users' data in 155 responses. Config said ~1M context, KV c…(opens the original post in a new tab)
9. 09:13
   X@nusawaqa· Discussion

   Their write-up, which is pretty open about it: https://blog.cloudflare.com/kv-cache-exposure-remediation/(opens the original post in a new tab)
10. 09:13
    X@nusawaqa· Discussion

    A bug in Cloudflare's AI service mixed bits of other people's prompts into 155 replies this summer. Only on huge prompt…(opens the original post in a new tab)
11. 04:20
    X@contentbuffer· Discussion

    Cloudflare disclosed that a config mismatch + SGLang bug in Workers AI (GLM-5.3) served some customers other customers'…(opens the original post in a new tab)
12. 03:50
    X@BlueRockID· Discussion

    Cloudflare disclosed a bug: 155 Workers AI responses to very long prompts contained other customers' data (Aug 27–Sep 9…(opens the original post in a new tab)
13. 02:08
    X@SaFeRamirez· Discussion

    Cloudflare probó su IA con prompts largos y todo dio 200 OK. Igual, 155 respuestas salieron con datos de otros clientes…(opens the original post in a new tab)
14. 01:22
    X@Dr1tz· Discussion

    @TheHackersNews Cloudflare's own write-up adds scope: one model (GLM 5.3), 155 responses, only prompts over 681,280 tok…(opens the original post in a new tab)
15. 01:21
    X@nurhayatdev· Discussion

    ⚠ Cloudflare disclosed a cross-customer data exposure in Workers AI’s GLM-5.3 serving path. From Aug 27–Sep 9, 155 re…(opens the original post in a new tab)
16. 01:12
    X@qihanblmg· Discussion

    There is also an infrastructure bug: Cloudflare disclosed a caching bug that caused other customers' data to be mixed into 155 GLM-5.3 responses. Note that this is not model hallucination; it is the cache layer inserting someone else's data into your response. Model problems and infrastructure problems are fixed in completely different ways: the former by adjusting the model, the latter by fixing…(opens the original post in a new tab)
17. 01:07
    X@joespano_· Discussion

    Cloudflare's hosted GLM-5.3 sent 155 responses with other customers' data mixed in. No hacker. Just a long prompt and a…(opens the original post in a new tab)
18. 01:00
    X@leetllm· Discussion

    The endpoint returned HTTP 200 with other customers' data. Cloudflare reports 155 affected GLM-5.3 Workers AI responses…(opens the original post in a new tab)

1. 22:17
   X@KeyserlingRoger· Discussion

   Cloudflare says a bug put other customers' data into 155 responses from its hosted GLM-5.3 model, Aug 27 to Sep 9. It's…(opens the original post in a new tab)
2. 21:26
   X@johnmorganFL· Discussion

   @TheHackersNews Credit to Cloudflare for patching fast and publishing the details. This is the AI bug class to watch: a…(opens the original post in a new tab)
3. 21:12
   X@viccy2025· Discussion

   Workers AI GLM-5.3: 155 responses leaked other tenants' data (KV cache). Who: prompts >681,280 tokens on @cf/zai-or…(opens the original post in a new tab)
4. 19:50
   X@ainewshublive· Discussion

   Cloudflare disclosed a cross-request KV cache exposure in its GLM-5.3 Workers AI model, tied to 155 responses. A cache…(opens the original post in a new tab)
5. 19:25
   X@Xray7K· Discussion

   Cloudflare's own write-up: only prompts over 681K tokens were hit, the model was pulled and fixed Sep 9, and it urges o…(opens the original post in a new tab)
6. 19:24
   X@Xray7K· Discussion

   Cloudflare's AI served one customer's data to another. Cloudflare says 155 GLM 5.3 responses held data from other cust…(opens the original post in a new tab)
7. 19:09
   X@itsMikeKipruto· Discussion

   Cloudflare's AI exposed customer data across user sessions due to a cache isolation flaw. 155 responses reportedly con…(opens the original post in a new tab)
8. 19:06
   X@yasinilhan· Discussion

   Cloudflare's new postmortem is a wild read. GLM-5.3 advertised a 1M-token context, but the GPU cache only held 681K. P…(opens the original post in a new tab)
9. 18:44
   X@TheHackersNews· Discussion

   🚨 Cloudflare's AI leaked customer data across user sessions. → 155 responses exposed other customers' data. → SGLang K…(opens the original post in a new tab)
10. 18:05
    X@STomoya_todo· Discussion

    On Cloudflare's AI infrastructure, "other people's input" was mixed into 155 responses. The cause was a combination of insufficient KV cache capacity and a bug where rejection processing returned HTTP 200. https://itpostjp.com/security/clou…(opens the original post in a new tab)
11. 17:00
    X@DanielShibz· Discussion

    On October 26, Cloudflare published an incident postmortem: from August 27 to September 9, 155 replies from GLM-5.3 on Workers AI mixed in content from requests other customers had just sent. The trigger condition was a prompt exceeding 681,280 t...(opens the original post in a new tab)
12. 14:36
    X@hkmrcash· Discussion

    Cloudflare publicly releases Workers AI platform incident report; 155 replies leaked data. AI SI and U October 28, 2026|Openai|G... https://youtu.be/kKUkonTN9sI?si=qU431KM2lai…(opens the original post in a new tab)
13. 13:21
    X@shailgn· Discussion

    GLM-5.3 on Cloudflare's own Workers AI was found to leak KV cache across requests. It advertises a 1,048,576 token context, but the prefill KV pool holds only about 681,280. For prompts exceeding this length, decoding reads other…(opens the original post in a new tab)
14. 12:57
    X@sxkshii_26· Discussion

    A Cloudflare/SGLang bug leaked other users' data into 155 GLM-5.3 responses. Prompts over 681,280 tokens were rejected…(opens the original post in a new tab)
15. 12:33
    X@thesystemprompt· Discussion

    Cloudflare's A I cache leaked other customers' prompts. Cloudflare discloses that a KV cache bug in its GLM-5.3 Workers…(opens the original post in a new tab)
16. 12:24
    X@SinghalPratul· Discussion

    14. ⚠ Cloudflare leaked other customers' data in 155 replies A GLM-5.3 cache bug on Workers AI, Aug 27 to Sep 9 Long p…(opens the original post in a new tab)
17. 12:10
    iThome (Taiwan)Selected

    Cloudflare AI service cache error: 155 GLM-5.3 responses mixed in others' data

    Cloudflare published the findings of its investigation into the data leak at its Workers AI: between August 27 and September 9, the GLM-5.3 model had a total of 155 responses that mixed in other customers' data; the fix was completed on September 9; other models were unaffected, and no evidence of deliberate exploitation was found. The incident stemmed from Cloudflare's deployment configuration and a program vulnerability in the open-source inference software SGLang: the GLM-5.3 input limit was set to about 1.05 million text units, but the GPU cache handling the input could hold only about 680,000; when input exceeded the cache capacity, SGLang did not reject the request as designed, causing responses to mix in content left in the KV cache by other customers. Gadi Evron, a researcher at the security firm Knostic, reported it on September 8; Cloudflare reproduced the problem the next day with two accounts, suspended the model service, and restored it the same day after the fix. Cloudflare acknowledged that before launch it knew the cache capacity did not match the input limit, but at the time it only checked response status and performance; SGLang users had reported the related vulnerability as early as July.

18. 12:00
    WebAI Threat Watch (aithreat.io)· Media coverage

    Penny for Your Prompts: A Measurement Study of Contextual Leakage in AI Chat Platforms(opens the original post in a new tab)
19. 11:39
    X@__kokumoto· Discussion

    Cross-tenant information leakage that occurred in GLM-5.3 on Cloudflare Workers AI. Due to an inconsistency between SGLang's KV cache management and the memory capacity settings, another user's request content was mixed into responses. 155 responses were affected between 8/27 and 9/9. 9…(opens the original post in a new tab)
20. 11:25
    X@alexlee20122022· Discussion

    On October 26, Cloudflare disclosed a leak in GLM-5.3, a model running on its Workers AI platform. The incident itself…(opens the original post in a new tab)
21. 11:02
    X@David888111888· Discussion

    2/5 📌 Multi-tenant isolation crisis: cross-request KV cache leak incident in Cloudflare Workers AI. Cloudflare issued a security advisory revealing a cross-request data leak vulnerability in the GLM-5.3 model running on Workers AI. Some overly long requests read...(opens the original post in a new tab)
22. 10:59
    X@doctorsab0· Discussion

    LLM serving is a shared system. Cloudflare's GLM-5.3 returned 155 responses with other customers' data: a 1M-token cont…(opens the original post in a new tab)
23. 09:20
    X@khomkrit· Discussion

    Cloudflare ออกมาเล่าว่าโมเดล GLM-5.3 บน Workers AI เคยมีบั๊กที่ทำให้ KV cache ข้ามไปโผล่ในคำขอของคนอื่นได้ และอธิบายว่า…(opens the original post in a new tab)
24. 08:44
    X@superdoccimo· Discussion

    An incident occurred on Cloudflare Workers AI in which another customer's data was mixed into responses. The impact was 155 responses, and it has already been fixed. The cause was a mismatch between the GPU's KV cache capacity and the input limit, along with a defect in the refusal handling. Even when HTTP 200 is returned, correctly…(opens the original post in a new tab)
25. 06:32
    X@Melmelmel69· Discussion

    For two weeks in late August and early September, Cloudflare’s GLM-5.3 model on Workers AI had a quiet leak. 155 respo…(opens the original post in a new tab)
26. 06:18
    X@Macfeecn· Discussion

    Cloudflare sold GLM-5.3 as 1M context, but its KV cache only fit 681,280 tokens. Prompts past that read other users' st…(opens the original post in a new tab)
27. 06:18
    X@tontontonton· Discussion

    https://gigazine.net/news/20261028-kv-cache-exposure-remediation/ #cloudnews #クラウド #AI Cloudflare Workers AIでホストされているGL…(opens the original post in a new tab)
28. 06:08
    X@__su888· Discussion

    GLM 5.3のKVプール(681,280)が文脈長未満で、SGLangが拒否要求を中断せず古いKVを読み155件の応答に他顧客データが混入。プールを約2倍にし受付で拒否を確定 / Cross-request KV cache expos…(opens the original post in a new tab)
29. 05:54
    X@SharedSapience· Discussion

    Cloudflare discloses GLM 5.3 data exposure, fixed in open-source SGLang Sources: https://blog.cloudflare.com/kv-cache-…(opens the original post in a new tab)
30. 05:35
    X@dev2saints· Discussion

    cloudflare's writeup on the glm-5.3 kv cache leak is how a postmortem should read. 155 affected responses, the exact ro…(opens the original post in a new tab)
31. 05:33
    X@kenebeii· Discussion

    📰 Gigazine: Cloudflare explains the KV cache leak between requests in the GLM-5.3 Workers AI model and SGLang's countermeasures https://gigazine.net/news/20261028-kv-ca…(opens the original post in a new tab)
32. 05:01
    X@RohanBhanotAI· Discussion

    cloudflare's report: https://blog.cloudflare.com/kv-cache-exposure-remediation/(opens the original post in a new tab)
33. 05:00
    X@arasorsdev· Discussion

    3/ Cloudflare says no other Workers AI model was affected and found no evidence of deliberate exploitation. The new dis…(opens the original post in a new tab)
34. 05:00
    X@arasorsdev· Discussion

    2/ UPDATE: Cloudflare’s Oct 26 report covers a GLM-5.3 Workers AI incident from Aug 27–Sep 9. It says 155 long-prompt r…(opens the original post in a new tab)
35. 05:00
    X@gigazine· Discussion

    Cloudflare explains KV cache leakage between requests in the GLM-5.3 Workers AI model and the countermeasures by SGLang(opens the original post in a new tab)
36. 05:00
    WebGigazine· Media coverage

    Cloudflare explains KV cache leakage between requests in the GLM-5.3 Workers AI model and the countermeasures by SGLang(opens the original post in a new tab)
37. 04:45
    X@tsnmediahub· Discussion

    For anyone buying hosted AI, the lesson is that tenant isolation in a model service is not only about accounts and perm…(opens the original post in a new tab)
38. 04:45
    X@tsnmediahub· Discussion

    ☁ Cloudflare says a bug in its GLM-5.3 model on Workers AI put fragments of other customers' requests into 155 replies…(opens the original post in a new tab)
39. 03:09
    X@LowdCode· Discussion

    Source: Cloudflare blog, "kv-cache-exposure-remediation": https://blog.cloudflare.com/kv-cache-exposure-remediation/(opens the original post in a new tab)
40. 03:04
    X@LowdCode· Discussion

    Cloudflare says 155 GLM 5.3 responses on Workers AI leaked other customers' data. Cause: the model accepted prompts up…(opens the original post in a new tab)
41. 02:36
    X@codigomaiz· Discussion

    🔐 Cloudflare reveló que su modelo GLM-5.3 en Workers AI mezcló datos de otros usuarios en 155 respuestas (27 ago–9 sep)…(opens the original post in a new tab)
42. 02:30
    X@KalinJunYAI· Discussion

    cloudflare sold a 1M context on GLM-5.3 when the cache only fit 681,280 tokens. i'd rather pay more for a smaller conte…(opens the original post in a new tab)
43. 02:01
    X@Karthikvarmamkv· Discussion

    the @Cloudflare glm-5.3 leak threw no errors. every response was a 200 🧯 - context: 1,048,576 tokens - kv cache pool:…(opens the original post in a new tab)
44. 02:00
    X@ArcyDev· Discussion

    A long prompt should never come with someone else’s data. Cloudflare’s incident report is a reminder to test tenant is…(opens the original post in a new tab)
45. 01:20
    X@andrewoshei· Discussion

    Cloudflare reports a tenant isolation bug in the GLM-5.3 Workers AI model leaked other users' data in 155 responses. Th…(opens the original post in a new tab)
46. 01:07
    X@danialhasan· Discussion

    @threepointone https://blog.cloudflare.com/kv-cache-exposure-remediation/(opens the original post in a new tab)
47. 01:04
    X@svallejo_tech· Discussion

    Fuente: https://blog.cloudflare.com/kv-cache-exposure-remediation/(opens the original post in a new tab)
48. 01:00
    X@Gael_DDOS· Discussion

    Cloudflare's GLM-5.3 leak: 155 responses read other tenants' KV pages ⚠ - 1M context, 681k prefill pool - a one-token…(opens the original post in a new tab)
49. 01:00
    X@petrusenko_max· Discussion

    Cloudflare says its Workers AI GLM-5.3 model exposed cross-request KV cache data, affecting 155 responses. The company…(opens the original post in a new tab)
50. 00:42
    X@themuscleai· Discussion

    155 responses leaked other customers' data through a one-token placeholder. Your inference cache is a tenant boundary,…(opens the original post in a new tab)
51. 00:40
    X@audiee_ai· Discussion

    https://blog.cloudflare.com/kv-cache-exposure-remediation/(opens the original post in a new tab)
52. 00:39
    X@audiee_ai· Discussion

    cloudflare's postmortem: glm-5.3 on workers ai sent some users text from other customers' prompts, 155 responses, per c…(opens the original post in a new tab)
53. 00:30
    X@BalmikeKvijay· Discussion

    Cloudflare found a cross-request KV cache exposure in a GLM-5.3 model on Workers AI: 155 responses affected. My read: t…(opens the original post in a new tab)
54. 00:09
    X@larsnow· Discussion

    https://blog.cloudflare.com/kv-cache-exposure-remediation(opens the original post in a new tab)
55. 00:08
    X@CEOScottFord· Discussion

    Cloudflare's GLM-5.3 bug sent pieces of other customers' prompts back in 155 responses. Credit to them for publishing t…(opens the original post in a new tab)
56. 00:00
    X@marioscheliga· Discussion

    GLM-5.3 on Workers AI put other customers' prompts into 155 responses (27 Aug to 9 Sep). 1M context, KV pool sized for…(opens the original post in a new tab)

1. 23:19
   X@dwkim601· Discussion

   Source: https://blog.cloudflare.com/kv-cache-exposure-remediation/(opens the original post in a new tab)
2. 23:17
   X@dwkim601· Discussion

   Cloudflare says a misconfiguration and an SGLang bug made its GLM-5.3 model on Workers AI return other customers' data…(opens the original post in a new tab)
3. 23:15
   X@daayushh· Discussion

   Cloudflare's GLM-5.3 leak shows what "HTTP 200" can hide. The model took 1,048,576-token prompts, but its KV cache poo…(opens the original post in a new tab)
4. 23:06
   X@AyushSarode07· Discussion

   A deep dive into how a KV cache bug exposed data across users and how Cloudflare fixed it. https://blog.cloudflare.com…(opens the original post in a new tab)
5. 23:04
   X@amir_mahla· Discussion

   cloudflare's postmortem: https://blog.cloudflare.com/kv-cache-exposure-remediation/sglang's side + operator guidance: h…(opens the original post in a new tab)
6. 23:01
   X@larsnow· Discussion

   Cloudflare disclosed that 155 responses on Workers AI contained other customers' data between August 27 and September 9…(opens the original post in a new tab)
7. 22:46
   X@DailyDarkWeb· Discussion

   ☁🚨 CLOUDFLARE: WORKERS AI RESPONSES EXPOSED OTHER CUSTOMERS' DATA Cloudflare says 155 responses from its GLM-5.3 Work…(opens the original post in a new tab)
8. 22:35
   WebAI Threat Watch (aithreat.io)· Media coverage

   Cloudflare GLM-5.3 Leak: 155 Responses Exposed [2026](opens the original post in a new tab)
9. 22:31
   X@farseen· Discussion

   https://blog.cloudflare.com/kv-cache-exposure-remediation/(opens the original post in a new tab)
10. 22:31
    X@farseen· Discussion

    cloudflare's glm-5.3 on workers ai sent 155 responses with other customers' data in them . every one was a prompt over…(opens the original post in a new tab)
11. 22:30
    X@andamonium_dev· Discussion

    1/ Cloudflare disclosed cross-request KV cache exposure in a Workers AI model, affecting 155 responses shared inferenc…(opens the original post in a new tab)
12. 22:19
    X@ryanpon· Discussion

    155 responses from GLM-5.3 on Cloudflare Workers AI held other customers' data over 13 days, all from prompts past 681,…(opens the original post in a new tab)
13. 21:57
    X@AgenticCoreAI· Discussion

    10. ☁ Cloudflare Bug Leaks GLM-5.3 Responses KV cache isolation flaw exposed other users’ data in 155 long-context re…(opens the original post in a new tab)
14. 21:51
    X@DeepWater_SI· Discussion

    Cloudflare's GLM-5.3 leak: the context window its memory couldn't hold For two weeks this summer, a very long prompt s…(opens the original post in a new tab)
15. 21:45
    X@heydeharro· Discussion

    Cloudflare: 155 Workers AI responses leaked other customers' data through a KV cache bug. Shared inference is the new…(opens the original post in a new tab)
16. 21:44
    X@Archive2050· Discussion

    Source: Cloudflare Blog, 2026-10-26 — Cross-request KV cache exposure in GLM-5.3 Workers AI (coordinated with SGLang /…(opens the original post in a new tab)
17. 21:44
    X@Archive2050· Discussion

    A long prompt on Workers AI came back with someone else's session still in the reply. On October 26, Cloudflare publis…(opens the original post in a new tab)
18. 21:12
    X@shailgn· Discussion

    GLM-5.3 on Cloudflare's own Workers AI was found to leak KV cache across requests. It advertises a 1,048,576 token context, but the prefill KV pool holds only about 681,280. For prompts exceeding this length, decoding reads other…(opens the original post in a new tab)
19. 20:41
    X@SohaibAnwa11· Discussion

    Sources: https://blog.cloudflare.com/kv-cache-exposure-remediation/ Upstream account: https://www.lmsys.org/blog/2026-1…(opens the original post in a new tab)
20. 20:37
    X@SohaibAnwa11· Discussion

    HTTP 200 isn't a tenant-isolation test. Cloudflare disclosed a fixed GLM-5.3 bug: rejected long prompts could reach st…(opens the original post in a new tab)
21. 20:34
    X@chataclaw· Discussion

    Cloudflare disclosed details of an incident in Workers AI's GLM-5.3 in which other people's prompts were mixed into responses. Although the KV pool was smaller than the context window, requests that did not fit did not error out but instead read old cache. An HTTP-code-only det...(opens the original post in a new tab)
22. 20:31
    X@evolvehqio· Discussion

    Cloudflare disclosed that GLM-5.3 on Workers AI returned other tenants' data in 155 responses, all on prompts over 681,…(opens the original post in a new tab)
23. 20:31
    X@7zSushi· Discussion

    ⚠ Cloudflare disclosed a cross-customer leak (Oct 26): from Aug 27 to Sep 9, 155 GLM 5.3 responses on Workers AI conta…(opens the original post in a new tab)
24. 20:29
    X@8zero7· Discussion

    Cloudflare's report: https://blog.cloudflare.com/kv-cache-exposure-remediation/(opens the original post in a new tab)
25. 20:27
    X@bitsquin· Discussion

    Cloudflare disclosed a security incident where 155 responses on its Workers AI deployment of GLM-5.3 inadvertently expo…(opens the original post in a new tab)
26. 20:25
    X@djcroman· Discussion

    Cloudflare's hosted GLM 5.3 sent 155 replies containing other customers' data. Prompts over 681,280 tokens didn't fit…(opens the original post in a new tab)
27. 20:09
    X@zgoodz· Discussion

    SGLang and Cloudflare disclosed an inference incident: from 8/27 to 9/9, GLM-5.3 on Workers AI, under PD-disaggregated deployment, could have ultra-long requests read the contents of other people's sessions. Root cause: the configured context was 1M tokens, but pre…(opens the original post in a new tab)
28. 09:00
    KnosticSelected

    Hallucination or cross-tenant: A large language model on Cloudflare Workers AI answers with another user's context

    Knostic researcher Gadi Evron found that the @cf/zai-org/glm-5.3 endpoint of Cloudflare Workers AI returns the contents of other users' inference sessions rather than a response to the submitted prompt when a request exceeds about 700,000 tokens. The leaked content included personal names, company names, invoice numbers, euro amounts, payment dates, source code information, and Agent task records. The author used black-box falsification to rule out hallucination and training data reproduction: the unique marker in the request disappeared, latency dropped sharply at about 700,000 tokens, the output had zero overlap with the input, and a shell call returned in one response matched a public repository that was created after the model's training cutoff and had a signed verification commit less than 4 minutes after the affected inference. The issue was reported via HackerOne on September 8, 2026; Cloudflare confirmed and deployed a fix on September 24, publicly disclosed it on October 26, and its blog analyzed in depth the upstream vulnerability in SGLang.

29. 07:10
    Unite.AISelected

    Cloudflare released a report stating that a GLM-5.3 flaw caused cross-tenant session leakage of 155 responses

    Cloudflare released an incident report on October 26, 2026: the GLM-5.3 model on its Workers AI platform suffered a cross-request KV cache leak, and 155 responses were mixed with data from other customers' sessions. The issue was reported by Knostic researcher Gadi Evron via HackerOne on September 8, 2026, and affected the service provided as @cf/zai-org/glm-5.3 since the model went live on August 27; Cloudflare fixed it and restored the model on September 9, stating that no evidence of deliberate abuse was found and that other Workers AI models were unaffected. Cloudflare has expanded the prefill KV pool to approximately 2.04 million tokens and deployed an engine patch; the relevant fixes have been merged into SGLang v0.5.20 through v0.5.22, and the company recommends that all SGLang operators upgrade to v0.5.22 or higher.

30. 06:01
    CloudflareSelected

    Cloudflare discloses cross-request KV cache leak in the GLM-5.3 model on Workers AI and coordinated fix with SGLang

    Cloudflare 披露，其 Workers AI 平台上 Z.ai 的 GLM-5.3 模型曾发生跨请求 KV cache 泄露，2026 年 8 月 27 日至 9 月 9 日间共 155 条回复包含其他用户请求的数据。受影响回复全部来自超过 681,280 token 的长提示，无证据显示漏洞被刻意利用，其他 Workers AI 模型未受影响。根因是两个问题叠加：prefill 节点的 KV cache 池仅 681,280 token，小于模型 1,048,576 token 的上下文窗口，而推理引擎 SGLang 在分离式 prefill/decode 模式下会把被拒的超长请求替换为单 token 占位符后照常传输，decode 节点因此读取了先前请求遗留的缓存页。

1. Time not given
   LMSYS OrgSelected

   SGLang security update: improved request isolation

   SGLang 维护方与 Cloudflare 协调披露 SGLang PD 分离服务中的跨请求 KV 缓存隔离缺陷。Cloudflare 的调查发现，其 Workers AI 上的 GLM-5.3 部署在 2026 年 8 月 27 日至 9 月 9 日间，超出 prefill 缓存容量的请求可能返回其他用户会话的内容，起因是配置的上下文长度 1,048,576 token 超过 prefill 缓存池的 681,280 token，被拒请求仍进入分离交接，decode 读到未为本请求填充的缓存页。相关的一类竞态是请求中止后已发出的 RDMA 写入落入被重新分配的槽位，造成无报错的单向跨请求泄漏，长上下文会扩大窗口；SGLang 自 v0.4.x 至 v0.5.21（MORI 后端至 v0.5.22）的 PD 默认配置受影响。维护方建议 PD 部署升级并运行能配对中止的路由器。

Current attention 64·Peak in the comparable period 85 (Oct 27 22:00)·Change over the comparable past 24 hours -21%