According to CoinMeta, Cloudflare disclosed a data leakage incident involving AI. A vulnerability in user isolation was found in the GLM-5.3 model on its Workers AI platform, which resulted in 155 responses being mixed with content previously submitted by other customers. The issue stemmed from GPU cache management; Cloudflare allowed the model to access approximately 1.05 million token of context, but the GPU cache responsible for processing inputs could only hold about 680,000 token. Such excessively long requests should have triggered an error, but the underlying inference engine SGLang mistakenly continued to execute them, allowing the model to read the cached data left by other users. The actual leakage was fixed on September 9th, and it was confirmed that other Workers AI models were not affected by this issue, nor was any evidence of malicious exploitation found.