Cross-request KV Cache Exposure in the GLM-5.3 Workers AI Model
On this page
On 26 October 2026, Cloudflare disclosed a security issue in its Workers AI inference platform: in the GLM-5.3 model, key–value (KV) cache state from one request could be exposed in subsequent requests. The company reported 155 affected responses and published a remediation post.
KV caches are the internal attention state that a transformer carries as it generates tokens. Normally each request has an isolated cache; an exposure means data from one user’s generation could leak into another’s, a serious confidentiality problem for a hosted model service. Cloudflare’s post describes how the issue was found, the affected window, and the remediation steps taken.
Why It Matters
This incident is a concrete example of a class of failure that is easy to miss when models are served as opaque APIs: the stateful internals of inference (caches, context windows, and serving infrastructure) can leak across users. It raises questions about isolation guarantees in hosted AI, about auditing and disclosure practices, and about whether “stateless” API boundaries are really secure. See Inference and Serving, Interpretability and Transparency, AI Safety, and AI in National Security.
Sources
- Cloudflare. Cross-request KV cache exposure in the GLM-5.3 Workers AI model (blog, 26 Oct 2026). https://blog.cloudflare.com/kv-cache-exposure-remediation/
- Hacker News discussion. https://news.ycombinator.com/item?id=50252159
Back to Recent Topics.