All stories
Cloudflare GLM-5.3 KV exposure Debates Source-backed

Cloudflare discloses GLM-5.3 cache exposure affecting 155 responses

The October 26 account describes an August-to-September incident fixed September 9, not a newly active breach.

Sources (1)

Why now October 26 disclosure; August 27 to September 9 exposure. Backfilled October 27 after a publication gap; this site's publication time is not the event date.

Cloudflare disclosed on October 26 that a request-handling failure in its Workers AI GLM-5.3 deployment exposed other customers’ data in 155 responses. According to the company’s account, the exposure window ran from August 27 to September 9. Cloudflare says it disabled the model, applied mitigations and restored it on September 9.

The incident involved the @cf/zai-org/glm-5.3 model and very long prompts. Cloudflare explains that the advertised context capacity exceeded the available prefill cache allocation. A rejected request could still be handed to the decoding stage, which then attended to stale cache pages containing data from another request.

The company says its investigation did not find evidence of deliberate exploitation and that no other Workers AI model was affected. Those scope and intent assessments are Cloudflare’s findings, not an independent forensic audit.

The disclosure is relevant beyond one provider because it shows how an inference system’s admission and cache handling can become a data-isolation issue. Cloudflare describes increasing the cache pool and fixing rejected-request handling, and recommends SGLang 0.5.22 or later. The new event is the public disclosure and technical explanation. Readers should not interpret the October 26 publication date as the start of an unresolved breach.