Fetching from the wire…
Public story · 2026-08-24 · high
A new method sent 1.1 million requests across 39 reseller endpoints and found more than a third of pairs share an upstream provider.
Why now: This detection method is new, and it gives anyone buying access through a reseller a way to check what's underneath before an outage forces the question.
CacheTracer sent 1.1 million requests across 636 endpoint pairs on 39 reseller endpoints, finding that 37.1% share a hidden upstream through prefix-cache reuse, per the CacheTracer paper.
That matters for anyone buying model access through a discount reseller. One cache reach sat nested inside at least 31 other nodes. A problem at one hidden upstream can ripple through resellers that look like separate vendors on their pricing pages.
The detection method, called Flood-then-Prove, sends a prompt through one endpoint to populate its cache. Then it sends the same prefix through a second endpoint to check whether the response comes back faster. A faster response means the two endpoints share cache state and therefore a backend. The researchers mapped a containment order seven layers deep across the full test set.
The 37.1% figure is a floor. CacheTracer only catches sharing when a shared cache returns a hit, so an upstream that skips prefix caching looks independent even when it's the same provider underneath. Anyone routing production traffic through the cheapest available endpoint is buying into a backend they can't name, with a blast radius they can't size.
Each link below shares sources, entities, or timing with this story.
Shared entity: LLM API / Same source domain / Earlier coverage
Both cover LLM API; reported by the same outlet (arxiv.org); earlier LLM API coverage from 2026-07-27.
Same source domain / Shared topic
Reported by the same outlet (arxiv.org); overlapping topics (cache, found).
Reported by the same outlet (arxiv.org); overlapping topics (cache, deep).
Reported by the same outlet (arxiv.org); overlapping topics (blast, cache).
Reported by the same outlet (arxiv.org); overlapping topics (channel, through).
Shared entity: Routing / Earlier coverage
Both cover Routing; earlier Routing coverage from 2026-08-17.
Shared entity: LLM API / Earlier coverage
Both cover LLM API; earlier LLM API coverage from 2026-08-16.
Both cover LLM API; earlier LLM API coverage from 2026-08-13.