173 points56 comments
Cloudflare Workers AI leaked other users’ text in GLM-5.3
What is Cross-request KV cache exposure in the GLM-5.3 Workers AI model (155 responses)?
A configuration mistake let the GLM-5.3 model reuse its KV cache across different requests, so some users saw fragments of other customers’ prompts in the model’s replies. The bug affected 155 long-context responses between Aug 27 and Sep 9 and has now been patched.
Who it is for
If you run or rely on Cloudflare’s Workers AI, verify you’re using the updated model version and watch for any anomalous output.