Cloudflare: GLM-5.3 KV Cache Mismatch Leaked Customer Data

Cloudflare published an incident report on October 26, 2026, revealing that between August 27 and September 9, the GLM-5.3 model on Workers AI produced 155 responses that contained fragments of other customers' requests. According to the report, none of the requests involved any hacking — the incident stemmed from two failures happening at the same time.

The first is a capacity mismatch. GLM-5.3 advertises a context window of roughly 1M tokens, but the GPU KV cache pool on the serving infrastructure holds only 681,280 tokens. Requests exceeding that ceiling are therefore marked for rejection, because there is not enough room for their KV cache. The second failure is that this rejection was never enforced: over-long requests were flagged as rejected but never actually terminated, so they continued to execute on the GPUs. Once those requests ran, fragments of other customers' requests — prompts longer than 681,280 tokens belonging to different tenants — ended up mixed into GLM-5.3's outputs.

The incident was surfaced in a third-party write-up published on October 31, 2026, which characterized the underlying bug as more instructive than the headline numbers. As of that report, Cloudflare had disclosed the mechanism through its postmortem, but no further details were available on how the affected customers were notified or what follow-up fixes were shipped. The affected window ran for roughly two weeks before ending on September 9.