Research
KVGuard: A KV Cache Safety Checker for LLM Serving Systems
Publication date unknown · Discovered arxiv.org
Page published
Publication date unknown · First observed: 5 Nov 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
KVGuard addresses silent output corruption and broken request isolation in LLM serving systems that conventional GPU memory sanitizers cannot detect.
KVGuard is a research prototype that detects KV-cache memory-safety violations in vLLM and SGLang using per-block permission and initialization metadata. The paper reports coverage of 29 historical bugs and discovery of eight previously unknown safety violations, with no false positives. Reported geometric-mean time-per-output-token overhead is 3.0% on vLLM and 7.4% on SGLang.