🔍 Read the full analysis: Could Your AI Session Be Affected By A KV Cache Leak? on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Cloudflare says a configuration mismatch and an SGLang bug caused its hosted GLM-5.3 model to return content from other customers’ requests in 155 responses between Aug. 27 and Sept. 9, 2026. The company took the model offline and restored a fixed version; it says it found no evidence of deliberate exploitation, but its report does not explain how affected customers were notified.
Cloudflare disclosed that its hosted GLM-5.3 model returned text from other customers’ requests in 155 responses between Aug. 27 and Sept. 9, 2026. The company attributed the leak to a mismatch in its memory configuration combined with a bug in the open-source SGLang inference engine; it says it took the model offline and found no evidence of deliberate exploitation.
Cloudflare’s incident report, published Oct. 26, says the affected responses came from requests containing more than 681,280 tokens. That limit was below the model endpoint’s advertised context window of 1,048,576 tokens. The issue involved the model’s key-value cache, or KV cache, which holds intermediate data generated while processing a prompt.
Cloudflare said its prefill GPUs had insufficient KV cache capacity for prompts beyond the lower threshold. A bug in SGLang then allowed a request rejected by one part of the system to pass through separate queues in a split prefill-and-decode setup. The decode stage read cache pages that had not been overwritten, exposing text from other users’ prompts and partial generations. The model returned an HTTP 200 response rather than an error.
Security researcher Gadi Evron of Knostic reported the issue through Cloudflare’s HackerOne program on Sept. 8. Cloudflare’s timeline says it began verification the following morning, declared an incident at 13:22 UTC and took the model offline 38 minutes later. It restored a fixed version that evening. The report says the 155 responses were recorded during the 14-day incident window, excluding Cloudflare’s own tests and the researcher’s requests when describing ordinary customer activity.
Infographic 01 · Cross-request KV cache exposure
How a long prompt read someone else's data.
A memory pool smaller than the advertised context window, plus an engine bug that let rejected requests through, made one hosted model attend over other customers' leftover pages.
The mismatch
tokens; Cloudflare knew before launch and expected an error
Only prompts longer than 681,280 tokens could reach the faulty path.
The eight pages
Cloudflare’s own example of one over-long request
written by this requeststill holding earlier users’ prompts and replies
- Decode reserves 8 pages for the long prompt.
- Prefill rejects it and swaps in a one-token placeholder…
- …but the rejected request still passes the handoff checks.
- Prefill writes one page; decode reads all eight.
- The client gets HTTP 200 and a coherent reply, partly built on someone else’s data.
Why pages hold old data
speed trade-off
- GPU memory is a shared pool of fixed-size pages
- Pages are not zeroed between requests; the next one overwrites
- Bookkeeping is the only wall between customers
The fix
9 Sep, 19:40 UTC
- Pool resized so every valid prompt fits
- Engine retires rejected requests before any memory transfer
- Upstream: SGLang v0.5.20 to v0.5.22; operators should run v0.5.22+
Still open
not in the report
- Whether and how affected customers were notified
- How many customers’ requests contributed the leaked data
- What kind of data it was
Sources: Cloudflare incident report (26 Oct 2026); SGLang/LMSYS (26 Oct 2026). Figures and the no-exploitation finding are Cloudflare’s own.
Powered by Thorsten Meyer AI
How the Cache Leak Affected Customers
The incident shows how a memory-management failure in an inference service can expose one customer’s input to another, even when the underlying language model is not itself responsible. For people using hosted AI with long prompts, that can include source code, business documents or other material submitted for processing. The report does not establish what specific kinds of customer information appeared in the affected responses.
Cloudflare said the requests were consistent in timing and volume with ordinary long-context use, rather than repeated attempts to exploit the flaw. It reported no evidence of deliberate exploitation. That distinction matters, but it does not remove the privacy implications: a response can appear coherent and still contain information from another session. Cloudflare said one observable clue was that affected answers arrived faster than expected.
The incident also underlines the limits of testing only for service availability. Cloudflare said it had tested HTTP response codes and performance at long context lengths, but had not checked the output’s content and coherence. In a system that handles prompts for multiple customers, validating only that a request returns successfully may not reveal whether the returned text belongs to the right user.
Products Worth Considering

Apple MacBook Air 15-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Liquid Retina Display, and Touch ID (Midnight)

Samsung 27-Inch Odyssey OLED G5 QHD Gaming Monitor (G50SF)

Logitech MX Ergo S Advanced Wireless Trackball Mouse USB-C – Graphite
As an affiliate, we earn on qualifying purchases.
Why KV Cache Pages Matter
A KV cache stores intermediate calculations a language model uses as it reads a prompt and generates a response. Reusing those calculations avoids repeating work, which can improve speed. Inference systems divide GPU memory into pages that requests borrow and return. According to Cloudflare, those pages are not cleared between requests; the next request is expected to overwrite them, a practice the company described as common and needed for performance.
That design depends on the system tracking page ownership correctly. In Cloudflare’s setup, one GPU pool processed prompts and built the cache, while another generated responses using that cache. The configuration mismatch meant the first stage could not write all the pages reserved for an oversized request. The SGLang bug allowed the second stage to proceed anyway, leaving some pages with earlier users’ data.
Cloudflare’s report also says the underlying SGLang issue had been reported publicly as issue #30233 on July 6, 2026. The company said it became aware of that report during its investigation. The disclosure comes amid wider concerns about security reporting volume: the source material also describes a separate Jump Crypto audit contest where AI-generated or AI-assisted submissions complicated review. That is a distinct issue from the Cloudflare leak and does not establish a connection between the two events.
““pages are not currently zeroed between requests; the next request that gets a page simply overwrites it.””
— Cloudflare incident report
Questions About Exposure and Notice
Cloudflare’s report gives the count and date range for affected responses, but the supplied material does not identify which customers were affected, what types of information appeared, or how many distinct customer accounts received exposed text. It also does not say how or when the company notified affected customers. The report’s finding that activity resembled ordinary use is Cloudflare’s assessment, not independent confirmation of every request’s origin.
The available account does not provide a full technical description of the fixed version, independent verification of the repair, or details on whether other deployments using the same SGLang behavior faced similar exposure. It is also unclear from the supplied report excerpt whether all potentially affected outputs were identified or whether customers can determine if their own sessions were involved.
Further Disclosure and Remediation
Cloudflare has said it restored a fixed version of the model after taking the endpoint offline. The next developments to watch are any additional technical detail from Cloudflare or SGLang, and whether Cloudflare explains its customer-notification process or provides affected users with a way to check their exposure. The source material does not give a date for further updates.
For users of hosted AI services, the report offers a specific reason to ask providers how they isolate cached request data, test long-context behavior and handle incident notifications. Those questions do not establish that other services have the same flaw; the confirmed disclosure here concerns Cloudflare’s GLM-5.3 endpoint and the incident period identified in its report.
Key Questions
What happened in the Cloudflare AI incident?
Cloudflare said its hosted GLM-5.3 model returned text from other customers’ requests in 155 responses between Aug. 27 and Sept. 9, 2026. The company attributed the incident to a configuration mismatch and a bug in SGLang.
What is a KV cache?
A key-value cache stores intermediate information a model generates while processing a prompt. It lets the system reuse that work when generating a response, but the incident showed that incorrect handling of cache pages can expose data between requests.
Was the leak deliberately exploited?
Cloudflare said it found no evidence of deliberate exploitation. Its report described the timing and volume of affected requests as consistent with ordinary long-context use. That is the company’s assessment of the incident.
What did Cloudflare do after the report?
Cloudflare said it took the model offline 38 minutes after declaring an incident and restored a fixed version that evening. A researcher had reported the issue through HackerOne on Sept. 8.
Were affected customers notified?
The supplied incident report does not say how affected customers were informed or identify which customers received another user’s text. Those details remain unclear from the available material.
Source: ThorstenMeyerAI.com
