GPU VulnDB

Database/AI/ML frameworks & serving

vLLM: race condition in single_type_kv_cache_manager.py common prefix handling

CVSS 1.3CVE-2026-129356AI/ML frameworks & servingcurated

Impact

vLLM 0.30.0 and 0.31.0 have a race condition in the common prefix handler of the KV cache manager (single_type_kv_cache_manager.py). The record does not say what an attacker gains. The CVSS vector rates confidentiality, integrity and availability impact all as low, and limited to the vLLM service itself. On a shared inference endpoint the KV prefix cache is used by requests from many clients, but the record does not say whether one client's data can reach another. The record says a public exploit exists and that exploitation is highly complex.

Who can reach it

Remote client of the vLLM API with low privileges (the vector says some authentication is needed, PR:L). Attack complexity is high.

What to do

Upgrade vLLM to 0.32.0, which carries fix commit e687a251bb056fa527846633f7a1b0a6c0f42374, then restart the serving processes. Restarting means reloading model weights onto the GPUs, so roll it through replicas behind the load balancer.

References

Related entries

All AI/ML frameworks & serving entries

This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.