Cloudflare says other customers’ data appeared in GLM-5.3 replies between August 27 and September 9, 2026. It says it corrected the memory configuration and patched the software before restoring access on September 9.
Cloudflare, the sole source for this account, said in its October 26, 2026 disclosure that 155 GLM-5.3 responses on its superintelligence (SI) hosting service contained data from other customers’ recent requests. The exposure lasted from August 27 to September 9, 2026, when it restored the model after repairs.
GLM is a text-generation model from Z.ai. Cloudflare launched version 5.3 on the service on August 27 with the fault already present. The company said no other hosted model was affected. The total of 155 counts responses, not customers; Cloudflare has not given the number of affected customers.
Cloudflare said it found no evidence of deliberate exploitation. After excluding its own tests and the researcher’s tests, the timing and volume of the identified requests were consistent with ordinary use of long prompts.
Cloudflare attributed the exposure to two overlapping problems. The first was a mismatch between the model’s context window—the amount of text it could accept—and the memory available to process it. The configured limit was 1,048,576 tokens, the small units into which a model divides text, but the prompt-processing cache could hold only 681,280. All 155 affected responses followed prompts longer than that smaller limit.
The service uses SGLang, open-source software for running models, with separate groups of graphics processors handling prompts and generating replies. The first group creates a key-value, or KV, cache: stored calculations that the second group uses to generate text. Cache memory is reused between requests rather than cleared, so the software must ensure each request reads only its own data.
The second problem was how SGLang handled a prompt that exceeded the cache capacity. Cloudflare said the software marked the request as rejected and replaced its prompt with a one-token placeholder, but failed to stop processing. It transferred the placeholder’s cache and signalled completion. The reply-generating processors then read memory that had not been overwritten, including earlier users’ prompts and partially generated replies.
The W27 keeps the Galaxy Z Fold 8 Ultra’s core hardware, with dark-red or dark-blue finishes, golden accents and 16GB of working memory in both storage versions.
2 min
Cloudflare acknowledged that it knew about the capacity mismatch before launch but expected oversized requests to return errors. Its tests checked successful response codes and performance, not whether the generated text was coherent. That gap let the fault pass its checks.
Security researcher Gadi Evron from Knostic alerted Cloudflare on September 8. The company said it reproduced the fault on September 9 and took the model offline within 40 minutes of declaring an incident.
Before restoring access that day, Cloudflare enlarged the cache so every prompt within the configured context window would fit. It also patched SGLang to stop rejected requests and return their original error before allocating memory or transferring cache data. It tested both changes against the researcher’s reproduction script and its own tests for detecting data crossing between requests.
Cloudflare said it applied the engine fix and additional safeguards across its SGLang deployments. Deployment checks now compare the running engine’s cache capacity with the configured context window rather than relying on successful response codes.
Further work includes a draft proposal to clear cache memory when it is released and broader support for checks that detect corrupted cache data. For other operators using separate prompt-processing and reply-generating systems, Cloudflare recommends measuring cache capacity after configuration changes and upgrading to SGLang 0.5.22 or later. Version 0.5.22 was released on October 19.
The satellite operator expects access to more than twice OneWeb’s current bandwidth in return for spending in 2027–2034, with initial IRIS² services expected from 2030.
3 min
The black, gold and green edition follows September’s sold-out preorders. Nintendo’s replenished launch-day stock also quickly sold out.
1 min
