New Technique Medium TD-2026-0032 For AI Eng, AppSec, Red Team AI-assisted
KVGuard: arXiv paper proposes safety checker for KV cache in vLLM and SGLang
The paper identifies a class of memory safety bugs in PagedAttention KV cache management, including double-free, use-after-free, out-of-bounds access, uninitialized read, and permission violation. It presents KVGuard, which uses shadow blocks and five safety checks to detect these violations. The authors report that a prototype covers 29 historical bugs in vLLM and SGLang and found 8 previously unknown violations.
Why it matters
These bugs can silently corrupt model outputs or break isolation between requests without crashing the server. Operators of multi-tenant LLM serving stacks may see wrong or leaked outputs that standard GPU sanitizers do not catch.
Do this
- Track this paper and review whether KVGuard's approach or code is available before evaluating it for internal LLM serving stacks.
- Review vLLM and SGLang patch histories for KV cache memory safety fixes and confirm deployed versions include them.
- For multi-tenant deployments, test request isolation under concurrent load to check for cross-request output leakage.
- Consider adding KV cache block-level invariant checks in staging environments, since NVIDIA Compute Sanitizer does not reason about block tables.
Signals
- NoCISA KEV
- 23/100Priority
- 60%Confidence
Evidence
“this block-based design exposes a previously unstudied class of memory safety bugs in the KV cache, including double-free, use-after-free, out-of-bounds access, uninitialized read, and permission violation”
“A prototype of KVGuard on vLLM and SGLang covers all 29 historical real-world KV cache safety bugs collected from the two systems' patch histories.”
“KVGuard further found 8 previously unknown safety violations while reporting no false positives.”
Sources
Affected / related
PagedAttention, vLLM, SGLang, KVGuard, NVIDIA Compute Sanitizer
Published 2026-11-04 08:22 UTC · updated 2026-11-08 17:52 UTC · AI-assisted summary, checked against quoted sources · How we score