← October 28, 2026 briefing · added in the 16:00 KST update
KV cache mixing between requests on Cloudflare Workers AI's GLM-5.3 exposed other customers' prompts and outputs in 155 responses (community report)
Verified (fact)
According to a GeekNews summary, responses from the GLM-5.3 model served on Cloudflare Workers AI contained other customers' prompts and generated content. The issue lasted from August 27 to September 9, 2026, and affected 155 responses. Other models reportedly did not have the same problem. The cause is described as KV cache exposure between requests.
Why it matters (interpretation)
This is a real-world case showing that reusing or sharing KV caches for serving optimization can break data isolation between customers in multi-tenant environments. Companies using managed inference APIs should separately verify the isolation guarantees of the serving layer.
Confidence medium · official source pending
Sources
- Community GLM-5.3 Workers AI 모델에서 요청 간 KV 캐시 노출, 응답 155건 영향 GeekNews
More about Cloudflare
-
Cloudflare open-sources 'Clef', a decision model for AI agents (report)
According to an InfoQ China headline, Cloudflare has open-sourced 'Clef', a decision model aimed at AI agents. Details such as model...
-
Cloudflare launches a CLI built for AI agents and plans to phase out its existing developer CLI, Wrangler (report)
InfoQ China reported that Cloudflare has released a command-line tool (CLI) designed for use by AI agents, and plans to gradually end...
-
vLLM Semantic Router team open-sources 'Decision 3.0,' multimodal decision models that also analyze images (report)
The vLLM Semantic Router team open-sourced 'Decision 3.0,' a family of multimodal decision models that analyze images directly, as well...
-
Cloudflare to acquire Deno, boosting self-hosted workerd; Deno runtime support ends in one year (community report)
According to a write-up by Simon Willison, Cloudflare is acquiring Deno. The goal is to make self-hosting workerd an officially...