Post

Breaking the Semantic Cache: A LiteLLM Case Study (CVE-2026-89032)

CVE-2026-89032 is a semantic cache isolation failure in LiteLLM that breaks API key-based user isolation. Unlike web caches that use deterministic keys, semantic caches match queries via fuzzy vector similarity with a configurable threshold; attackers can craft queries semantically similar to other users’ cached responses to trigger cache hits and leak confidential information. The vulnerability enables multiple attack classes: information disclosure across user boundaries, LLM hallucinations, and arbitrary tool execution. The root cause stems from the inherent fuzzy-matching nature of semantic caching combined with deployment pressure to lower similarity thresholds for cost efficiency. NVISO discovered and responsibly disclosed the vulnerability; a patch was released in LiteLLM v1.101.0-rc.1.

Read original article

This post is licensed under CC BY 4.0 by the author.