🔥 Trending on HN

A shared AI memory returned someone else’s data: Cloudflare’s GLM-5.3 disclosure

2 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
KV cache

Temporary information kept so the model can answer faster.

tenant isolation

A rule that keeps one customer’s data separate from another customer’s data.

SGLang

Open-source software used to run AI models on servers.

What happened

On October 26, 2026, Cloudflare published a report about a security incident in Workers AI. The affected model was GLM-5.3, a model from Z.ai. From August 27 through September 9, some customers received replies containing material from other customers’ recent requests. Cloudflare counted 155 affected responses. Every affected response involved a prompt longer than 681,280 tokens.

Cloudflare says the requests matched ordinary long-context use after its own tests and the researcher’s requests were removed. It found no evidence that someone repeatedly exploited the issue. Gadi Evron of Knostic reported the problem through HackerOne on September 8. Cloudflare reproduced it the next morning, took the model offline within 40 minutes of declaring an incident, and restored it later that day. The company’s account is documented in Cloudflare’s report.

The background

AI systems keep temporary information so they can process long prompts faster. This is called a KV cache. In a cloud service, many customers share computing hardware, so tenant isolation must remain intact.

Cloudflare says GLM-5.3 had a context limit of 1,048,576 tokens. Its prefill cache pool held only 681,280 tokens. The company knew about this mismatch before launch. Its testing checked successful HTTP responses and performance, but did not sufficiently check the returned text and its coherence. A configuration mismatch combined with an error-handling bug in SGLang, the inference software Cloudflare used. Some requests that should have stopped instead moved through the system incorrectly, creating a path for old data from another request to appear in a reply.

Why it matters

The incident shows that AI safety is not only about the model’s behavior. The serving system, memory management, and failure paths also protect customer data. A successful status code does not prove that requests stayed isolated.

The number of affected responses was limited, but the trust issue is serious. Business prompts can contain confidential work or personal information. Cloudflare says no other Workers AI model was affected. Still, the case shows why shared-memory AI systems need tests that inspect outputs, not only speed and availability.

What the report confirms

After disabling the model, Cloudflare expanded the cache pool to about 2.04 million tokens. That is roughly twice the model’s context limit. It also patched SGLang so rejected requests stop before moving into the next processing stage. Cloudflare applied the new checks and hardening across its production SGLang deployments. It recommends SGLang version 0.5.22 or later for operators.

Separately, the Hacker News post received 100 points and 32 comments. That reaction measures community attention. It does not prove that the original report is correct.

What remains unknown

The public report does not identify the affected customers or describe the exact data seen in individual replies. The 155 figure counts responses, not confirmed people or accounts. The report also describes Cloudflare’s deployment. It is not an audit of every service that uses SGLang.

What to watch next

Operators should compare the model’s context limit with its actual cache capacity after changes to models, quantization, or hardware layout. They should update the inference engine and test returned text during failure cases. Cloudflare is also working on stronger cache-ownership checks and formal verification. The model was fixed, but keeping customers isolated remains an ongoing engineering duty.

💬 Debating the GLM-5.3 Workers AI KV-cache exposure

HN commenters largely frame the issue as a failure of tenant isolation, Cloudflare configuration, SGLang behavior, and test design rather than a defect in the GLM-5.3 model itself. The thread also disputes how responsibility and the article title should be understood.

  • The central security concern is cross-tenant isolation: multiple users may share GPU resources and KV-cache storage, so a separation failure could cause data from one session to overlap with another. Some commenters recommend non-shared hosts for highly sensitive enterprise workloads.
  • Responsibility is disputed. One commenter says the primary problems are Cloudflare configuration and an SGLang bug, not the model. A commenter presenting themselves as being from Cloudflare says GLM-5.3 was named to identify the affected endpoint, not to claim that the model itself was defective. Others still find the title misleading.
  • The comments discuss article figures of a 1,048,576-token context window and a 681,280-token KV-cache pool on the prefill GPUs. Some ask whether the advertised window was practically usable; another interpretation is that 681,280 is only a single-prefill limit, while multi-turn sessions with cache hits could still reach one million tokens.
  • Because long-context tests returned responses instead of errors, one commenter suggests that leaked or incorrect output may have been counted as a successful test. This is a commenter’s interpretation, not an independently verified result.
  • The root-cause analysis was praised for clearly explaining why testing missed the problem, and for quickly taking the model out of service and remediating the issue.
  • One user self-reported that, during the past year, ChatGPT dictation sometimes became what seemed like a different conversation after processing. This anecdote does not establish a link to the incident.

initial digest at 32 comments (revision 1). We fetched 32 comments and sampled 32 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

An AI reply included another customer’s data

📰 Full story: A shared AI memory returned someone else’s data: Cloudflare’s GLM-5.3 disclosure

Cloudflare found a bug that mixed customer information in one AI service.

1 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
KV cache

A temporary information area that helps AI answer faster.

tenant isolation

Keeping each customer’s information separate.

Hacker News

A website where technology stories receive community attention.

💡 The gist

  • GLM-5.3 sometimes returned another customer’s data.
  • The problem lasted from August 27 to September 9.
  • Cloudflare fixed the setup and the software.

Cloudflare runs Workers AI, a service for using AI models. GLM-5.3 is a text model from Z.ai. It uses SGLang, open-source software for running models.

AI systems save temporary information to answer faster. This saved area is called a KV cache. Different customers must still remain separate. That separation is called tenant isolation.

The model accepted very long prompts. The available cache space was smaller than its context limit. An error path failed to stop some requests correctly. Some replies then included data from earlier requests.

Cloudflare says 155 responses were affected. All involved prompts longer than 681,280 tokens. The company found no evidence of deliberate exploitation. A researcher reported the issue through HackerOne. Cloudflare reproduced it the next morning. It took the model offline within 40 minutes. It fixed the problem and restored the model that day.

Cloudflare enlarged the cache pool to about 2.04 million tokens. It also changed SGLang to stop rejected requests earlier. Cloudflare says other Workers AI models were not affected.

The Hacker News post drew 100 points and 32 comments. Those numbers show attention, not correctness.

The deeper lesson is about careful testing. A fast, successful reply is not enough. AI operators must also check failed requests. They must ensure each customer’s information stays separate.

The public report does not name the customers involved. It does not describe the exact data in each reply. Operators using SGLang should update to version 0.5.22 or later.

💬 A possible mix-up in shared AI memory

The discussion focuses more on how shared computers are separated and tested than on the GLM-5.3 model itself.

  • When many users share a GPU, their temporary KV cache must stay separate. If it does not, one user’s information could appear in another session. Some commenters say sensitive work should use a private machine.
  • Some commenters blame Cloudflare’s setup and a problem in the SGLang engine, not GLM-5.3. A commenter presenting themselves as being from Cloudflare says the model name only identifies the affected service. Others say the title can still sound misleading.
  • The comments discuss a 1,048,576-token context window and a 681,280-token cache for the first long-reading step. One view is that the smaller number may limit only one initial load, so a long conversation could still reach the larger window.
  • A long-input test returned an answer instead of an error. One commenter thinks a leaked or wrong answer may have been counted as success.
  • People liked the clear explanation and the quick stop-and-fix response.
  • One user self-reported a strange ChatGPT dictation experience, but that does not prove it was caused by this incident.

initial digest at 32 comments (revision 1). We fetched 32 comments and sampled 32 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

The AI put one person’s words in another person’s answer

📰 Full story: A shared AI memory returned someone else’s data: Cloudflare’s GLM-5.3 disclosure

Cloudflare says one AI made a small mix-up.

1 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
KV cache

A little work shelf that helps AI answer faster.

Workers AI

A place where people can use AI.

GLM-5.3

The name of the AI that made the mix-up.

Cloudflare is a company that helps run AI. GLM-5.3 is the AI’s name. Workers AI is the place where it worked.

The AI keeps little bits of work nearby. That helps it answer faster. This little work shelf is called a KV cache. The shelf should stay with the right person. Sometimes, it did not. Some answers included another person’s words.

Cloudflare counted 155 mixed-up answers. Cloudflare found no sign someone did this on purpose. It turned the model off. It fixed the setup and the computer program. It brought the model back on September 9.

Cloudflare says other Workers AI models were not affected. Hacker News talked about this report. It had 100 points and 32 comments. Those numbers show attention. They do not prove the report is true.

People who run AI must keep everyone’s information apart.

💬 When shared AI memory gets mixed up

Some commenters think the problem was the divider between users’ shared computer spaces, not the AI’s brain.

  • Each user needs a separate little desk. If the desks mix, one person’s temporary AI notes might appear in another person’s work. Some commenters say important work should use a private computer.
  • Some people blame the service setup and the SGLang tool, not the GLM-5.3 model. Others say the title sounds as if it blames the model.
  • The service discussed 1,048,576 pieces of text, but one first-reading cache held 681,280. That smaller number might apply only to one first step. A strange answer may also have been mistaken for a successful test.
  • People liked the clear explanation and the quick fix. One user reported a strange ChatGPT dictation experience, but that story does not prove this incident caused it.

initial digest at 32 comments (revision 1). We fetched 32 comments and sampled 32 across the thread. These are HN users’ reports, not independently verified facts.

Sources