Cloudflare says 155 GLM-5.3 Workers AI replies mixed in other customers' data
Sorted AI29 OctCloudflare says that between 27 August and 9 September 2026, some customers using the GLM-5.3 model on Workers AI received responses that contained data from other customers' earlier requests to the same model. It reports 155 affected responses, all to prompts longer than 681,280 tokens, says it found no evidence of deliberate exploitation, and that no other Workers AI model was affected. Cloudflare says a security researcher reported the issue on 8 September; it took the model private the next day, fixed a KV-cache misconfiguration and an SGLang engine bug, and restored the model the same day. Treat the counts and timeline as Cloudflare's disclosure.
Why it matters: If you send long prompts or private text to a hosted model, ask the vendor how they isolate KV cache and tenant memory between requests. For sensitive work, prefer shorter contexts or providers that document isolation checks.
Cloudflare ↗AI-assisted, reviewed by Sorted AI
▲ 0
0 comments