Cloudflare details GLM-5.3 cache exposure on Workers AI
Cloudflare says 155 responses were affected during the GLM-5.3 cache incident and that SGLang was patched after the fault was traced.
Cloudflare published its explanation on October 26, 2026 of a cross-request KV cache exposure in the GLM-5.3 model (@cf/zai-org/glm-5.3) hosted on Workers AI. According to the company, 155 responses were affected in total, all of them for prompts exceeding 681,280 tokens.
Cloudflare investigators said the affected requests matched normal long-context usage patterns in timing and volume, concluding that deliberate exploitation was unlikely. The incident stemmed from a combination of a misconfiguration and a bug in the open-source inference engine SGLang.
GLM-5.3 was deployed with a context window of 1,048,576 tokens but a prefill KV cache pool that could hold only 681,280 tokens. A bug in how SGLang handled oversized requests in disaggregated prefill/decode mode replaced rejected requests with a one-token placeholder and forwarded placeholder KV cache to the decode node.
Remediation and advice for operators
Cloudflare disabled the model, adjusted the KV cache pool to hold about 2.04 million tokens, and patched SGLang so rejected requests are discarded and the number of transferred tokens is validated. The disclosure was reported by gigazine.net.
The company recommends that operators upgrade to SGLang v0.5.22 or later and enable delayed KV release, limit Mooncake transfer batches, and enable output corruption detection.
Quick answers
How many responses were affected by the GLM-5.3 cache exposure?
Cloudflare says 155 responses were affected, all of them for prompts exceeding 681,280 tokens.
Which SGLang version does Cloudflare recommend?
Cloudflare recommends upgrading to SGLang v0.5.22 or later.