Independent tech intelligence, checked against primary sources.

TechPulseMind Useful technology.
No manufactured hype.

Cloudflare fixes GLM-5.3 cache leak that exposed 155 Workers AI responses

Cloudflare says a GLM-5.3 bug on Workers AI exposed data across 155 responses to prompts over 681,280 tokens; the model was fixed and restored on September 9.

Cloudflare fixes GLM-5.3 cache leak that exposed 155 Workers AI responses

Cloudflare says a tenant isolation bug in the GLM-5.3 model on Workers AI caused 155 responses to include data from other customers’ sessions between August 27 and September 9, 2026. The company published its account on October 26 alongside a coordinated post from SGLang, the open-source inference engine at the center of the incident.

According to Cloudflare, only prompts longer than 681,280 tokens were affected, and no other model hosted on Workers AI was impacted.

Two overlapping problems

Cloudflare attributes the exposure to a misconfiguration on its side combined with a bug in SGLang. The company deployed GLM-5.3 (@cf/zai-org/glm-5.3) with a 1,048,576-token context window, but the KV cache pool on its prefill GPUs held only 681,280 tokens. Cloudflare says it knew about the mismatch before launch and expected the engine to return an error for prompts it could not process.

The second issue was how SGLang handled over-length requests in disaggregated prefill/decode mode. A prompt that did not fit was internally marked as rejected and swapped for a one-token placeholder, but prefill still generated a KV cache for the placeholder and sent it to the decode node. That node received a “transfer complete” message and assumed the previous cache had been overwritten with the full prompt, when most of the earlier state remained. Responses were then generated over stale cache pages holding data from earlier requests.

Timeline and remediation

Security researcher Gadi Evron of Knostic reported the issue through Cloudflare’s HackerOne program on September 8. Cloudflare reproduced it the next morning, took the model offline within 40 minutes of declaring an incident, deployed a corrected configuration and a patched inference engine, and restored service the same day.

To fix the configuration, Cloudflare aligned GLM-5.3 with its GLM-5.2 setup, freeing enough GPU memory for the prefill KV pool to hold about 2.04 million tokens — roughly twice the 1,048,576-token context length.

Cloudflare says the requests it identified, after excluding its own testing and the researcher’s, match ordinary long-context use in timing and volume. “We found no evidence of deliberate exploitation,” the company wrote.

The report is published as a coordinated disclosure with SGLang’s maintainers, and Cloudflare says it has hardened its production SGLang deployments and continues to work with upstream maintainers on this and similar bugs. The full write-up, including the response timeline, is available on the Cloudflare blog.

Related reading

Sources

Some links on this page may be affiliate links. If you buy through them we may earn a commission at no extra cost to you. See our affiliate disclosure.