Cloudflare has disclosed that a flaw on its Workers AI platform let 155 responses from the GLM-5.3 model contain data belonging to other customers' sessions. The company published its incident report on Monday 26 October, about seven weeks after it fixed the problem.
The exposure began on 27 August 2026, the day the model launched on the platform. A security researcher, Gadi Evron, reported the issue on 8 September through the company's bug bounty programme. Cloudflare took the model offline, fixed it and restored it on 9 September, and says no other model on the platform was affected.
Only very long prompts triggered the leak. Every affected response followed a prompt longer than 681,280 tokens. After excluding its own tests and the researcher's requests, Cloudflare says the remaining traffic matched ordinary long-context usage in timing and volume and shows no evidence of deliberate exploitation.
Two faults combined. Cloudflare had deployed the model with a context window of 1,048,576 tokens, but the memory pool on its prefill processors held only 681,280. The company knew of the mismatch before launch and expected over-length prompts to return errors, and its long-context tests checked response codes and speed but not the quality of the output. The second fault sat in the open-source SGLang inference engine: in a split prefill-and-decode setup, an over-length prompt was marked as rejected and replaced with a one-token placeholder, but the system still built cache for it and signalled completion. The decoding node then read stale cache pages left by earlier requests, and the client received a success code rather than an error. A user had reported the engine bug on 6 July.
On 9 September the company reproduced the leak from two accounts within 25 minutes, declared an incident at 13:22 UTC, set the model to private by 14:00 and restored it publicly at 19:40 with a corrected configuration and an engine patch of its own. The memory pool is now about 2.04 million tokens, roughly twice the context length.
The engine's maintainers merged fixes that reject such requests at admission, shipped in version 0.5.20 and refined in later releases; Cloudflare advises operators to move to 0.5.22 or later, since the fix was not backported. The maintainers measured throughput effects within two per cent. Cloudflare says it is working with them on checksums for cache transfers and plans to formally model the safety of its transfer protocol.
Operators running the engine in split mode are told to check that cache size matches context length on every node after any change to weights or memory settings, and to enable corruption detection where available.





