Research
CacheTracer Uses Prefix-Cache Reuse as a Side Channel and Finds 37.1% of LLM API Reseller Pairs Share Hidden Upstreams
CacheTracer (arXiv 2608.20732, Aug 21) sent 1.1 million API requests across 636 endpoint pairs on 39 LLM API reseller endpoints, using a Flood-then-Prove primitive that populates cache state through one endpoint and tests whether another can reuse it. It found 37.1% of measured pairs share cache reach, a containment order spanning seven layers, and one cache reach contained within at least 31 other nodes. For anyone routing production traffic through a cheap reseller, this is a measurable blast-radius argument: a single upstream confidentiality or integrity failure propagates to users of seemingly independent providers.
↳ Follow the thread