CacheTracer Exposes Hidden Dependencies in LLM API Reseller Chains

Researchers introduced CacheTracer, the first API-only measurement technique for uncovering hidden dependencies within the LLM API reseller ecosystem. Multi-level resale chains let user requests pass through undisclosed upstream providers that can inspect or alter prompts and responses, creating confidentiality and integrity risks that existing single-provider audits fail to capture.

CacheTracer exploits prefix-cache reuse as a side channel to infer cache-reach relationships between resellers. Its Flood primitive seeds fresh cache state at one endpoint, while Prove checks whether another endpoint can reuse that state, filtering out false positives from self-induced cache hits.

Testing 39 reseller endpoints with 1.1 million requests across 636 endpoint pairs, the team found shared cache reach in 37.1% of pairs, a dependency hierarchy spanning seven layers, and some cache states contained within at least 31 other nodes — with structures varying by model. The results, validated through consistency checks and controlled experiments, expose a concentrated, model-specific dependency graph where a single upstream failure could create a large blast radius across supposedly independent downstream resellers.