Cloudflare Discloses AI Model Bug That Leaked Other Customers' Data in 155 Responses
A long-prompt bug in Cloudflare's Workers AI exposed pieces of other customers' requests between August 27 and September 9, and has been fixed.
Cloudflare said in a blog post on its own site dated October 26 that its GLM-5.3 model on Workers AI sometimes returned data from other customers' requests. Workers AI is Cloudflare's service for running AI models for developers and businesses.
Cloudflare says two problems overlapped. The model's memory cache was smaller than the longest prompts it accepted, and a bug in SGLang, an open-source engine that runs AI models, let those oversized prompts continue instead of returning an error. The affected responses went to prompts longer than 681,280 tokens, and Cloudflare counts 155 of them.
The exposure began August 27, when the model launched. A security researcher, Gadi Evron of Knostic, reported it on September 8 through Cloudflare's HackerOne bug bounty program. Cloudflare took the model offline on September 9 and restored it the same day after the fix. It says the pattern looks like ordinary long-prompt use and it found no evidence of deliberate exploitation.
Cloudflare says it patched SGLang so oversized requests now fail with an error, and it added further safeguards to its production systems. It says no other Workers AI model was affected. The post gives no specific steps for customers of the service. Its advice is aimed at teams running SGLang themselves, who should check cache size against prompt length and upgrade to SGLang v0.5.22 or later.
If you or a vendor you rely on uses Workers AI, ask whether GLM-5.3 was used with very long prompts during that window. More broadly, avoid pasting customer data, passwords or contracts into AI tools that you have not vetted, and ask vendors how they separate one customer's data from another's.
- Cloudflare blog: KV cache exposure remediation (Cloudflare)
Is AI recommending your business?
Our AI visibility audit checks what ChatGPT, Gemini and Google's AI say about you, and includes a 1:1 review of the results.
See the audits