CVE-2026-129356: vLLM: race condition in KV cache prefix handler
MEDIUMA race condition in vLLM's Common Prefix Handler (single_type_kv_cache_manager.py) in versions 0.30.0 and 0.31.0 could let a low-privileged remote user trigger inconsistent KV cache state, with limited confidentiality, integrity and availability impact. Severity is medium (CVSS 5.0) because exploitation requires high attack complexity and the reporter itself describes it as difficult. The data does not show active exploitation: it is not in CISA KEV, no EPSS score is available, no Nuclei template exists and no SSVC decision is recorded, although VulDB states a public exploit exists, which conflicts with our own exploit data, so treat that claim as unverified. Teams running vLLM as shared multi-tenant inference infrastructure should prioritize it over single-user deployments, since cross-request cache confusion is the realistic risk. Upgrade to vLLM 0.32.0 (patch commit e687a251bb056fa527846633f7a1b0a6c0f42374) during the next scheduled maintenance window.
What is the risk?
Medium overall. Network-reachable, needs low privileges (an authenticated or API-keyed inference user) and no user interaction, but high attack complexity because the attacker must win a timing window in prefix-cache handling. Impact is rated low across confidentiality, integrity and availability. No KEV listing, no EPSS data and no scanner template; VulDB claims a public exploit, which is unconfirmed here. Exposure is highest for shared inference endpoints with prefix caching enabled and concurrent load. Exact consequences (for example cross-request data mixing versus instability) are not detailed in the advisory, so treat them as uncertain.
How does the attack unfold?
What systems are affected?
| Package | Ecosystem | Vulnerable Range | Patched |
|---|---|---|---|
| vLLM | pip | — | No patch |
Do you use vLLM? You're affected.
How severe is it?
What is the attack surface?
What should I do?
1 step1) Upgrade vLLM to 0.32.0 or later, which includes commit e687a251bb056fa527846633f7a1b0a6c0f42374 (PR 61126, issue 61673). 2) Inventory deployments to identify instances running 0.30.0 or 0.31.0, including container images and pinned requirements. 3) Until patched, consider disabling automatic prefix caching on shared or multi-tenant endpoints, accepting the performance cost. 4) Restrict inference API access to authenticated, trusted clients and apply rate limiting to reduce the chance of high-concurrency timing abuse. 5) Monitor for anomalous output inconsistencies, engine errors or crashes in the KV cache manager, and unusual bursts of concurrent requests with shared prefixes from a single client.
How is it classified?
Which compliance frameworks are affected?
This CVE is relevant to:
Frequently Asked Questions
What is CVE-2026-129356?
A race condition in vLLM's Common Prefix Handler (single_type_kv_cache_manager.py) in versions 0.30.0 and 0.31.0 could let a low-privileged remote user trigger inconsistent KV cache state, with limited confidentiality, integrity and availability impact. Severity is medium (CVSS 5.0) because exploitation requires high attack complexity and the reporter itself describes it as difficult. The data does not show active exploitation: it is not in CISA KEV, no EPSS score is available, no Nuclei template exists and no SSVC decision is recorded, although VulDB states a public exploit exists, which conflicts with our own exploit data, so treat that claim as unverified. Teams running vLLM as shared multi-tenant inference infrastructure should prioritize it over single-user deployments, since cross-request cache confusion is the realistic risk. Upgrade to vLLM 0.32.0 (patch commit e687a251bb056fa527846633f7a1b0a6c0f42374) during the next scheduled maintenance window.
Is CVE-2026-129356 actively exploited?
No confirmed active exploitation of CVE-2026-129356 has been reported, but organizations should still patch proactively.
How to fix CVE-2026-129356?
1) Upgrade vLLM to 0.32.0 or later, which includes commit e687a251bb056fa527846633f7a1b0a6c0f42374 (PR 61126, issue 61673). 2) Inventory deployments to identify instances running 0.30.0 or 0.31.0, including container images and pinned requirements. 3) Until patched, consider disabling automatic prefix caching on shared or multi-tenant endpoints, accepting the performance cost. 4) Restrict inference API access to authenticated, trusted clients and apply rate limiting to reduce the chance of high-concurrency timing abuse. 5) Monitor for anomalous output inconsistencies, engine errors or crashes in the KV cache manager, and unusual bursts of concurrent requests with shared prefixes from a single client.
What systems are affected by CVE-2026-129356?
This vulnerability affects the following AI/ML architecture patterns: model serving, LLM inference servers, multi-tenant inference platforms, OpenAI-compatible API gateways.
What is the CVSS score for CVE-2026-129356?
CVE-2026-129356 has a CVSS v3.1 base score of 5.0 (MEDIUM). The EPSS exploitation probability is 0.30%.
What is the AI security impact?
Affected AI Architectures
MITRE ATLAS Techniques
AML.T0029Denial of AI ServiceAML.T0040AI Model Inference API AccessAML.T0049Exploit Public-Facing ApplicationCompliance Controls Affected
What are the technical details?
Original Advisory
A vulnerability was found in vllm-project vllm 0.30.0/0.31.0. This vulnerability affects unknown code of the file single_type_kv_cache_manager.py of the component Common Prefix Handler. The manipulation results in race condition. The attack may be performed from remote. Attacks of this nature are highly complex. It is stated that the exploitability is difficult. The exploit has been made public and could be used. Upgrading to version 0.32.0 is able to resolve this issue. The patch is identified as e687a251bb056fa527846633f7a1b0a6c0f42374. You should upgrade the affected component.
Exploitation Scenario
An attacker with a low-privilege API key to a shared vLLM endpoint sends many concurrent requests that share a common prompt prefix, trying to hit the timing window in the Common Prefix Handler. If the race is won, KV cache blocks may be handled inconsistently, causing another tenant's request to receive corrupted or mismatched context, or causing the engine to error or degrade. Because attack complexity is high, the attacker needs repeated attempts and knowledge of the serving configuration, which makes reliable exploitation difficult.
Weaknesses (CWE)
CWE-362Concurrent Execution using Shared Resource with Improper SynchronizationPrimaryCWE-362Concurrent Execution using Shared Resource with Improper SynchronizationCWE-362 — Concurrent Execution using Shared Resource with Improper Synchronization: The product contains a concurrent code sequence that requires temporary, exclusive access to a shared resource, but a timing window exists in which the shared resource can be modified by another code sequence operating concurrently.
- [Architecture and Design] In languages that support it, use synchronization primitives. Only wrap these around critical code to minimize the impact on performance.
- [Architecture and Design] Use thread-safe capabilities such as the data access abstraction in Spring.
Source: MITRE CWE corpus.
CVSS Vector
CVSS:3.1/AV:N/AC:H/PR:L/UI:N/S:U/C:L/I:L/A:LReferences
- github.com/vllm-project/vllm/
- github.com/vllm-project/vllm/commit/e687a251bb056fa527846633f7a1b0a6c0f42374
- github.com/vllm-project/vllm/issues/61673
- github.com/vllm-project/vllm/pull/61126
- github.com/vllm-project/vllm/releases/tag/v0.32.0
- vuldb.com/cve/CVE-2026-129356
- vuldb.com/submit/1043823
- vuldb.com/vuln/437419
- vuldb.com/vuln/437419/cti
Timeline
Related Vulnerabilities
CVE-2026-6173210.0Decepticon: ChatML token injection yields sandbox RCE
Same package: vllmCVE-2024-110419.8vllm: RCE via unsafe pickle deserialization in MessageQueue
Same package: vllmCVE-2026-259609.8vllm: SSRF allows internal network access
Same package: vllmCVE-2025-472779.8vLLM: RCE via exposed TCPStore in distributed inference
Same package: vllmCVE-2024-90539.8vllm: RCE via unsafe pickle deserialization in RPC server
Same package: vllm