● Platform: Cloudflare Workers AI
● Model affected: GLM-5.3
● Class: cross-request data exposure
● Detail published: limited beyond the disclosure
Cloudflare has disclosed a cross-request data exposure flaw in GLM-5.3 running on Workers AI - a class of bug where data submitted in one request can become visible in another.
On shared inference infrastructure this is the worst shape of vulnerability, because the data crossing the boundary is whatever customers sent to the model: prompts, documents, keys pasted into a context window. Details beyond the disclosure itself have not been published.
It is the third infrastructure-layer AI security disclosure this week, alongside Cohere's findings on Qwen3.8 and the Cathedral report on Anthropic's internal systems. None of the three is a model capability failure. All three are the plumbing.
Sources
Rio Rundown, 27 October.