Cross-request KV cache exposure in the GLM-5.3 Workers AI model (155 responses)

The issue involved the key-value (KV) cache used by GLM-5.3, a model available through Workers AI. In transformer models, this cache retains intermediate attention data so a model can generate subsequent tokens more efficiently. The exposure occurred across requests: cached state associated with one request could affect processing of another, creating a risk that a response might incorporate information from a different user’s prompt. The reported scope was 155 responses, and the problem concerned this model’s inference path rather than KV storage as a general customer-facing database.

Cloudflare addressed the flaw by changing the inference handling to keep cached state isolated between requests, preventing one request from reusing another request’s KV data. The incident highlights a security requirement specific to shared AI-serving infrastructure: performance optimizations such as cache reuse must preserve strict boundaries between users, even when requests run concurrently or use the same model instance. The important questions for readers are how the exposure was detected, what information could appear in affected responses, and how the remediation was validated; the supplied text identifies the model, cross-request failure mode, response count, and remediation focus.