Your AI Answer In A Stranger’s Session: Cloudflare’s KV Cache Leak And The AI Spam Drowning Bug Hunters
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: Your AI Answer In A Stranger’s Session: Cloudflare’s KV Cache Leak And The AI Spam Drowning Bug Hunters on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the latest gadgets delivered free with Prime

  • Fast, free delivery on millions of items
  • Prime Video, Amazon Music and more included
  • Member-only deals all year
Start your free Prime trial Free trial for eligible customers · Cancel anytime
As an affiliate, we earn on qualifying purchases.

TL;DR

Cloudflare’s October 26 incident report says its hosted GLM-5.3 model returned text from other customers’ requests in 155 responses between August 27 and September 9, 2026. The company attributes the issue to a context-window configuration mismatch combined with a bug in the SGLang inference engine; it says the model is fixed and it found no evidence of deliberate exploitation. A separate audit contest highlights another AI-related security pressure: the growing volume of reports that require human review.

Cloudflare says its hosted GLM-5.3 AI service returned text from other customers’ requests in 155 responses over a two-week period, after a configuration mismatch and an inference-engine bug allowed some oversized prompts to pass through incorrectly. The company says it took the affected model offline after reproducing the issue, restored a fixed version the same evening, and found no evidence of deliberate exploitation; the disclosure also leaves open how affected customers were notified.

Cloudflare’s incident report, published October 26, says the affected responses occurred from August 27 through September 9, 2026, on the Workers AI endpoint for Z.ai’s GLM-5.3 model. The report says every affected request involved a prompt longer than 681,280 tokens. After excluding Cloudflare’s own tests and the researcher’s requests, the company said the remaining activity was consistent in timing and volume with ordinary long-context use, rather than repeated attempts to exploit the flaw.

The company attributed the incident to two faults acting together. Cloudflare configured a 1,048,576-token context window, while the memory pool used to store a request’s intermediate model data could hold only 681,280 tokens. A bug in the open-source SGLang inference engine then allowed requests that should have been rejected to continue through a split prefill-and-decode system. Cloudflare says the result was that the decode stage could read memory pages not fully written for that request, including text from other users’ prompts and partial responses.

Security researcher Gadi Evron reported the issue through Cloudflare’s HackerOne program on September 8. Cloudflare’s report says the company began verification the next morning, declared an incident at 13:22 UTC and took the model offline 38 minutes after declaring the incident. It says a fixed version was restored that evening. The report does not specify how many customers were affected, what kinds of content appeared in the responses, or how Cloudflare informed those customers.

At a glance
reportWhen: Incident ran August 27–September 9, 202…
The developmentCloudflare disclosed that a bug in its GLM-5.3 service exposed other customers’ request text in 155 responses, while a separate audit contest illustrates the review burden from AI-generated bug reports.
Thorsten Meyer AI
Cloudflare Workers AI · GLM-5.3

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.

155responses that contained other customers' data
14 daysexposure window, 27 Aug to 9 Sep 2026
38 minfrom declaring the incident to taking the model offline
noneevidence of deliberate exploitation, per Cloudflare

The mismatch

tokens; Cloudflare knew before launch and expected an error

  • Context windowadvertised1,048,576
  • KV pool before the fixprefill GPUs681,280
  • KV pool after the fixabout twice the window~2.04 M

Only prompts longer than 681,280 tokens could reach the faulty path.

The eight pages

Cloudflare’s own example of one over-long request

P0written now
P1left over
P2left over
P3left over
P4left over
P5left over
P6left over
P7left over

written by this requeststill holding earlier users’ prompts and replies

  1. Decode reserves 8 pages for the long prompt.
  2. Prefill rejects it and swaps in a one-token placeholder…
  3. …but the rejected request still passes the handoff checks.
  4. Prefill writes one page; decode reads all eight.
  5. 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

Why Shared GPU Memory Matters

The incident shows how a boundary failure in an AI service can expose information even when the underlying model is not itself producing a deliberate disclosure. In this case, Cloudflare says GPU memory pages were reused between requests without being cleared, a performance practice that depends on the inference engine correctly managing which pages belong to each request. The two faults together broke that separation for some very long prompts.

For customers, the concern is not limited to whether an answer is coherent or whether a service returns an error. A response can look like a normal model output while containing material supplied by someone else. Organizations sending source code, internal documents or other sensitive text to hosted models may want to know which service endpoints were affected, what data could have been returned, and whether their account was involved. Cloudflare’s public report does not answer those customer-specific questions.

The report also connects to a different strain on security work: sorting credible vulnerabilities from a rising number of low-value submissions. Jump Crypto’s $1 million Firedancer audit contest drew 809 reports from 162 researchers and yielded 127 unique findings. Jump believes AI agents found or wrote nearly all the submissions, according to the source report. The contest and the Cloudflare incident are separate events, but together they illustrate two challenges: AI systems can create new failure modes, while AI-produced material can increase the work required to identify genuine security issues.

Products Worth Considering

Apple MacBook Air 15-inch Laptop with M5 Chip, 24GB Memory, 1TB SSD, Midnight

Apple MacBook Air 15-inch Laptop with M5 Chip, 24GB Memory, 1TB SSD, Midnight

Jackery Explorer 1000 v2 Portable Power Station, 1024Wh LiFePO4 Battery

Jackery Explorer 1000 v2 Portable Power Station, 1024Wh LiFePO4 Battery

Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation

Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation

As an affiliate, we earn on qualifying purchases.

How the Two System Faults Aligned

Large language models store intermediate calculations in what is known as a key-value cache, or KV cache, so they do not need to process the entire prompt again for every generated token. Cloudflare’s report says its inference system allocates this data in fixed-size memory pages. Those pages are not cleared between requests; instead, the next request assigned to a page is expected to overwrite its contents. That approach can improve performance, but it makes correct allocation and request handling essential.

Cloudflare ran the affected model in a disaggregated arrangement: one GPU pool processed prompts and built the KV cache, while another generated the answer using that cache. According to the report, an oversized prompt was marked as rejected by SGLang and replaced with a one-token placeholder. In the split setup, the queues checked whether the request had finished, but did not also check whether it had been rejected. The decode side had reserved space for the original prompt, so it could read pages that prefill had not overwritten. Cloudflare says the client received an HTTP 200 response, and the token count could match the long prompt, even though the response contained unrelated text.

Cloudflare says it had known about the difference between the advertised context limit and available memory before launch. It expected the inference engine to return an error for prompts that exceeded capacity, and said its long-context testing checked HTTP responses and engine performance without validating the model output and coherence. The company also said it learned during its investigation that a related SGLang bug had been reported as issue #30233 on July 6, 2026.

The second security-workflow story comes from Jump Crypto’s Firedancer audit contest. The 809 reports produced 127 unique findings, according to the source account, and Jump believes AI agents were responsible for or helped write nearly all of them. The report also cites other signs of pressure on review systems: curl ended its bug bounty in January, Bugcrowd began banning “submission farming” in March, and HackerOne recorded 46,947 submissions in March 2026. These figures have different sources and contexts; they do not establish that every submission was AI-generated or invalid.

“The affected requests are consistent in timing and volume with ordinary long-context use, not with someone repeatedly exploiting the bug.”

— Cloudflare’s incident report

Customer Impact Still Unspecified

Cloudflare’s report gives a count of 155 affected responses, but it does not state how many distinct customers or sessions were involved, or whether any particular category of sensitive data appeared. It also does not describe how customers were notified, whether the company contacted every potentially affected account, or what steps customers should take if they used the endpoint during the incident window.

The company says it found no evidence of deliberate exploitation and that the remaining requests appeared consistent with ordinary use. That is Cloudflare’s finding, not proof that every affected response has been identified or that no data was viewed or retained by a recipient. The report does not provide independent verification of the investigation’s scope. It also does not detail whether other models or configurations using SGLang were exposed to the same failure mode.

For the Firedancer contest, the source material reports Jump Crypto’s belief that AI agents found or wrote nearly every submission, but does not provide a breakdown showing which findings were AI-generated, how that was determined, or how many reports were duplicates or unusable. The raw counts describe that contest and its submissions; they do not establish a general rate of AI-generated security reports.

Follow-Up on Fixes and Notifications

Cloudflare says it restored a fixed version of GLM-5.3 on the evening of September 9 after taking the service offline during its response. The company’s published account establishes that the affected endpoint was returned to service, but does not lay out a further public timetable for customer notifications or additional technical disclosures. Readers will need to look for updates from Cloudflare on the scope of affected accounts, its notification process and any broader review of configurations that use the same inference engine.

For users of hosted AI services, the key follow-up is whether providers publish enough detail for customers to assess exposure: the affected dates and endpoints, the kind of information that could have been returned, and the actions taken to prevent recurrence. Cloudflare’s report identifies the configuration mismatch and SGLang behavior involved, but does not say whether the upstream project has released a fix or describe any coordinated changes across other deployments.

Jump Crypto’s new bounty platform, according to the source report, requires researchers to stake 1 SOL, valued there at about $100 per report. That rule is intended to make submissions costlier, but the source does not provide results showing whether it reduces low-quality reports or affects participation. The effectiveness of that approach, like the full customer impact of the Cloudflare incident, remains to be seen.

Key Questions

What data did Cloudflare’s AI service expose?

Cloudflare says 155 responses contained text from other customers’ requests. Its public report does not specify what kinds of prompts or information were involved, or how many distinct customers were affected.

Which model and dates were involved?

The issue affected Cloudflare Workers AI’s hosted GLM-5.3 endpoint, from August 27 through September 9, 2026, according to the incident report published October 26.

Was the GLM-5.3 model itself at fault?

Cloudflare attributed the incident to a memory-capacity configuration mismatch on its side and a request-handling bug in the SGLang inference engine. The report does not blame the GLM-5.3 model itself.

Did Cloudflare find evidence of an attacker exploiting the flaw?

Cloudflare says it found no evidence of deliberate exploitation. It said the remaining affected requests were consistent in timing and volume with ordinary long-context use after excluding its own tests and the researcher’s activity.

How were affected customers notified?

The incident report does not explain how customers were notified or identify the number of affected customers. Those details remain unresolved in the published account.

Source: ThorstenMeyerAI.com

You May Also Like

The Marshall Islands Pays Every Citizen A Basic Income. One Year In, What The First Nationwide UBI Can And Can’t Tell Us

The Marshall Islands completed Enra’s first year, paying $215.85 to 40,714 people in its final 2026 round. Its effects remain unevaluated.

PHP2Shell And A Node.js Security Release On The Same Day: What To Patch, Restart And Check Now

PHP fixed two critical RCE flaws and Node.js patched ten vulnerabilities on Oct. 21. Learn which systems need updates, redeploys or restarts.

Four AI Deals In One Day: Who Is Building The Plumbing Of The Agent Economy

Four developments on Oct. 29 put investment into AI infrastructure for voice, agent identity, building design and personalized shopping.

Freezers At 50 °C: The Second Wave Of Attacks On Europe’s Supermarket Refrigeration

Refrigeration controllers at more than 115 European stores were reportedly switched into defrost or heating mode overnight, disrupting freezers.