Cloudflare just released a blog on a cross-tenant data exposure vulnerability I reported to them. The vuln is in Cloudflare Workers AI’s GLM-5.3 provider. This was later also reported upstream to SGLang.
Cloudflare wrote a great post on the topic. While inference vulnerabilities are a class of their own, this one is certainly interesting! Large requests could return content from another customer’s inference session.
Check out their post here:
https://blog.cloudflare.com/kv-cache-exposure-remediation/
As a CEO I don’t often get to do the actual research with the Knostic team, and this was a pleasure.
A special note of thanks to the Cloudflare team, and specifically Avishek Sarkar, Kevin Flansburg, and Michelle Chen. Working with them has been a true pleasure, and I wish all bug bounty programs worked like theirs.
Cloudflare BlogCross-request KV cache exposure in the GLM-5.3 Workers AI model, and our remediation with SGLangA configuration mismatch in GLM 5.3 on Workers AI triggered an SGLang bug that exposed KV cache pages across requests. Here is how we addressed the issue.