Cloudflare disclosed on October 26 that a Workers AI failure caused 155 GLM-5.3 responses to include data from other customers’ requests. The company said the exposure affected requests between August 27 and September 9.
What went wrong
According to its incident report, the model accepted contexts larger than its configured key-value cache pool. An SGLang error-handling bug then allowed rejected requests to reach decoding and read stale cache data. Affected prompts exceeded 681,280 tokens.
Cloudflare said the problem was confined to @cf/zai-org/glm-5.3 and did not affect other Workers AI models. Researcher Gadi Evron of Knostic reported it on September 8.
Remediation and the remaining distinction
The model was disabled on September 9, within 40 minutes of the incident being declared, and restored after configuration and engine fixes. Cloudflare says it found no evidence of deliberate exploitation. That statement does not mean customer data was never exposed: the contaminated responses are the confirmed exposure.
Source
Cloudflare: KV cache exposure and remediation (October 26)
Feature image: AI-generated editorial illustration for Guava Insights. Conceptual artwork, not an actual product photograph, interface or facility.
