Cross-request KV cache exposure in the GLM-5.3 Workers AI model (155 responses)
On September 8, 2026, Cloudflare received notice from security researcher Gadi Evron about a problem in the GLM-5.3 Workers AI model. The issue, reported via Cloudflare's HackerOne program, caused responses to contain data from other customers' requests. Cloudflare worked quickly to identify and fix the problem, restoring service on September 9.
The issue stemmed from a misconfiguration of the open-source inference engine, SGLang, which had a 1,048,576-token context window but only had space for 681,280 tokens in its KV cache pool. Additionally, a bug in SGLang's handling of over-length requests led to stale data being included in responses. Cloudflare corrected the configuration by increasing the pool size, patched the engine to handle rejections safely, and proposed these changes upstream.
Written by urgent.news from Hacker News's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
