When AI Spam Buries The Signal For Security Researchers
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: When AI Spam Buries The Signal For Security Researchers on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get everyday essentials 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 reported that its hosted GLM-5.3 model returned other customers’ data in 155 responses between August 27 and September 9, 2026. Separately, Jump Crypto says nearly every report in a 809-submission Firedancer audit contest was found or written with AI agents, prompting a 1 SOL submission fee. The cases highlight a tension: AI can help find security flaws, but reviewing a flood of reports remains costly.

Cloudflare disclosed October 26 that its hosted GLM-5.3 model returned data from other customers’ requests in 155 responses over 14 days, while a separate security contest showed how AI-generated bug reports are straining review systems. The two developments expose different risks in AI security: failures in systems handling customer data and a growing volume of reports that researchers and platform operators must assess.

Cloudflare said the affected responses came from prompts longer than 681,280 tokens sent to its Workers AI endpoint for GLM-5.3, a text-generation model from Z.ai. Security researcher Gadi Evron of Knostic reported the issue through Cloudflare’s HackerOne program on September 8. Cloudflare’s incident report says it found 155 affected responses between August 27 and September 9.

The company traced the leak to two faults working together: a mismatch between the advertised 1,048,576-token context window and the smaller memory pool available to process prompts, and a bug in the open-source inference engine SGLang. Cloudflare says an oversized request was marked rejected but passed through queues in its split processing setup. The decode stage then read memory pages that had not been overwritten, exposing text from other requests.

Cloudflare says it took the model offline 38 minutes after declaring an incident and restored a fixed version that evening. It found no evidence of deliberate exploitation, and said the affected requests were consistent in timing and volume with ordinary long-context use. Separately, Jump Crypto’s $1 million Firedancer audit contest drew 809 submissions from 162 researchers, which the company said amounted to 127 unique findings. Jump believes AI agents found or wrote nearly every submission and has introduced a fee of 1 SOL per report.

At a glance
reportWhen: Cloudflare published its incident repor…
The developmentCloudflare disclosed a cross-customer data leak in its hosted AI service as a separate bug-bounty contest illustrated how AI-generated submissions can burden security reviewers.
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

Two Costs of AI-Assisted Security

The Cloudflare incident shows how a flaw in infrastructure around a model can expose information across customer sessions, even when the model itself is not identified as the source of the bug. Cloudflare’s report attributes the failure to its configuration and SGLang’s handling of rejected requests. The disclosure matters to customers using hosted AI services because it makes memory isolation and long-context testing concrete security concerns, not just performance questions.

The contest presents a different operational problem. AI tools may help researchers identify vulnerabilities or prepare reports, but each submission still takes time to validate. Jump’s fee shifts some of that cost back to submitters. That may deter low-quality reports, but the source material does not establish whether it will improve the accuracy of submissions or limit participation by independent researchers.

Together, the cases underline that AI can affect both sides of security work: the systems being tested and the tools used to test them. The consequences are distinct. One involved customer data appearing in responses; the other involved a high volume of contest reports, not evidence that the reports themselves caused a security breach.

Products Worth Considering

LG UltraGear 27-inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDR True Black 400

LG UltraGear 27-inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDR True Black 400

As an affiliate, we earn on qualifying purchases.

From Model Leak to Report Flood

Cloudflare’s system used a disaggregated design: one GPU pool processed prompts and created a key-value, or KV, cache, while another generated responses from that cached work. Cloudflare says cache pages were not cleared between requests; the next request was expected to overwrite them. In the incident, an oversized prompt was rejected by one part of the system but was not treated as rejected by both admission queues. The second stage read pages the first had not replaced.

Cloudflare said it had known of the context-window and memory-pool mismatch before launch, but expected requests that exceeded capacity to return an error. Its testing at long context lengths checked HTTP response codes and engine performance, not the model’s output and coherence. The SGLang issue involved had been reported upstream in July, according to Cloudflare, which said it noticed that report during its investigation.

The report-submission issue has surfaced elsewhere. The supplied source says curl ended its bug bounty in January 2026, Bugcrowd began banning “submission farming” in March, and HackerOne recorded 46,947 submissions in March 2026. Those are separate examples, not a single measured trend with a stated common baseline. They show why platforms are considering ways to manage volume, but do not by themselves establish how many submissions were AI-generated.

“Pages are not currently zeroed between requests; the next request that gets a page simply overwrites it.”

— Cloudflare

What the Reports Do Not Establish

Cloudflare’s report does not say how it notified affected customers, and the supplied material does not specify how many customers or distinct sessions were involved. The company said it found no evidence of deliberate exploitation, but that statement does not establish whether every exposed response was identified or whether recipients recognized the foreign text.

It is also unclear from the available account whether the SGLang bug has been fixed upstream or whether Cloudflare’s mitigation addresses all configurations that could trigger it. On the contest side, Jump’s view that AI agents were involved in nearly every report is not presented with a method for determining that, and the 127 unique findings figure does not show how many were valid, novel, or actionable.

Follow-Up on Fixes and Fees

Cloudflare says it restored a fixed version of the model after taking the endpoint offline, but the supplied incident account does not describe further independent testing or a customer-notification process. Those details would help clarify the scope of exposure and how the company plans to prevent similar failures in its inference stack. The status of the related SGLang issue is also a point to watch.

Jump’s new platform requires researchers to spend 1 SOL per report, about $100 according to CoinDesk. Future contest results may show whether the fee reduces duplicate or low-value submissions while keeping the program accessible to researchers. The source material does not provide results from the new policy, so its effects remain untested.

Key Questions

What data did the Cloudflare model leak?

Cloudflare said some responses contained text from other customers’ requests. Its report counted 155 affected responses, but the supplied material does not detail the content of each response or identify the customers.

How did the leak happen?

Cloudflare attributed it to a mismatch between the model’s advertised context window and available memory, combined with an SGLang bug. Rejected oversized requests passed through parts of the system, which then read memory pages that still contained data from other requests.

Did Cloudflare say the leak was deliberate?

No. Cloudflare said it found no evidence of deliberate exploitation and described the affected request pattern as consistent with ordinary long-context use.

Why is Jump Crypto charging for bug reports?

Jump introduced a 1 SOL fee per report after its Firedancer contest received 809 submissions that it said reduced to 127 unique findings. The company believes AI agents found or wrote nearly every report; whether the fee improves report quality has not yet been shown.

Source: ThorstenMeyerAI.com

You May Also Like

When AI Agents Begin Internal Permission Exchanges

Investigation reveals AI agents exchanging unauthorized messages and files, raising questions about control, authority, and safety in autonomous systems.

Futuristic Ventures: Integrating AI Security in Our Upcoming AI Projects

AIThis post was created with the assistance of artificial intelligence (AI). As…

Exploring the Human Impact of Ethical AI Security

AIThis post was created with the assistance of artificial intelligence (AI). We…

Uber Drivers’ Data At Risk? What To Do After The AI Breach

Uber says a breach at vendor Fireworks.ai exposed some Uber data. Letters identify affected information; reported notices include driver’s license data.