Cloudflare discloses repaired GLM 5.3 cross-request data exposure
Cloudflare has disclosed a repaired Workers AI incident in which GLM 5.3 responses could contain information from earlier users' requests. A prompt-cache capacity mismatch and an inference-engine rejection bug combined in the separated processing system. The company corrected the configuration and engine before restoring access in September, then published the coordinated disclosure on October 26.
Artificial Intelligence··Morning
The October disclosure concerns an earlier exposure
Cloudflare, the internet infrastructure company, disclosed a repaired data-exposure incident affecting GLM 5.3 on its Workers AI model-hosting service. Responses could contain information from earlier users' requests. The October 26 coordinated disclosure concerns an exposure between August 27 and September 9; the company says it restored access after correcting the issue in September. Its account identifies GLM 5.3 as the only affected model on the service.[1], [2]
A rejected request continued into response generation
Cloudflare says its prompt-processing cache pool was smaller than the advertised context window. In SGLang, the open-source inference engine, a rejection bug allowed an oversized request to continue through separated prompt-processing and response-generation machines. The response generator could then read cache pages containing state left by earlier requests. The cache holds intermediate information used by the model while it prepares its answer.[1]
Configuration and engine fixes preceded restored access
A researcher reported the issue through HackerOne, the vulnerability-reporting platform, on September 8. Cloudflare declared an incident the next day, restricted access, enlarged the cache pool and patched the engine before restoring public use. Follow-up work addressed memory release, transfers between machines and deployment checks comparing capacity with context length. Cloudflare recommends version 0.5.22 or later for operators of separated SGLang deployments.[1]
