Skip to content
See the World Through ScienceA project of ALLATRA

Cloudflare Says One of Its Hosted AI Models Returned Other People's Data

AI & Technology

Republish this story

Our work is licensed under Creative Commons BY-NC 4.0. You may republish this piece for free — with credit to ALLATRA Media and a link to the original, unedited beyond length trims, and not for commercial use.

Read the full license

Four computer memory modules stacked at an angle against a black background.
Memory modules of the kind stacked in server hardware, where space is reused from one request to the next (illustrative)."Electronic collection computer random access memory RAM modules on the black background" by driashkin, via Freepik, Freepik licence · Freepik-License

Cloudflare said 155 responses from the GLM-5.3 language model on its Workers AI service contained data belonging to other customers. The company's incident report, published Oct. 26, 2026, says the exposure was present from the model's launch on Aug. 27, 2026, until it was fixed on Sept. 9, 2026, and that no other model hosted on Workers AI was affected.

Cloudflare published the report as a coordinated disclosure alongside the maintainers of SGLang, the open-source inference engine it runs for the model. It closes with two instructions for other operators running the same setup: measure the cache pool against the advertised context length on every node, and upgrade to SGLang v0.5.22 or later.

Cloudflare traced the fault to two overlapping problems. It had deployed the model with a 1,048,576-token context window while the cache pool on the machines that process prompts held only 681,280 tokens. The company says it was aware of the mismatch before launch and had expected the engine to turn it into an error.

SGLang instead marked an over-length prompt as rejected, built a cache for a one-token stand-in, and passed it to the next stage as though the full prompt had been processed. The reply was then written over memory pages that still held data from earlier requests. "In total, 155 responses were affected, all to prompts longer than 681,280 tokens," the report says.

The report says GLM-5.3 was the only model affected because every other deployment of this kind on Workers AI has a cache pool larger than its context length. It also says the same bug had already been reported to SGLang by another user in July 2026.

Cloudflare says that after excluding its own testing and the researcher's, the affected requests match ordinary long-context use rather than repeated probing, and that it found no evidence of deliberate exploitation. That finding rests on the company's own logs. The report credits the discovery to Gadi Evron of Knostic, who reported it through Cloudflare's HackerOne bug bounty program.

Sources

Spot an error?

Spot an error?

Report an error

Spotted a mistake on this page? Tell us what's wrong and our editors will take a look.

What kind of problem?

Only if you'd like us to be able to follow up. We won't use it for anything else.

We correct mistakes openly. Select any text to flag it. Fixes are logged under our Corrections Policy.

Report an error

Reporting on

Cloudflare Says One of Its Hosted AI Models Returned Other People's Data

What kind of problem?

Only if you'd like us to be able to follow up. We won't use it for anything else.

We read every report. Corrections are logged publicly.