Cloudflare experienced a data leak between August and September 2026, where fragments of one tenant's KV cache state were exposed in 155 responses from its GLM-5.3 model on Workers AI. This occurred due to a combination of a configuration mismatch in the KV cache pool size and a bug in the SGLang library's disaggregated serving mode. The issue was resolved by increasing the KV pool capacity and patching SGLang to properly handle rejected requests, preventing stale data from being accessed. AI
IMPACT Highlights the critical need for robust KV cache management and SGLang's disaggregated serving to prevent tenant data exposure in LLM deployments.
RANK_REASON The article details a specific bug and its remediation in a deployed AI model, which falls under tooling and infrastructure rather than a new model release or significant industry shift.
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →