Inference corruption via uncleared recycled KV blocks
Published Apr 27, 2026 · Updated Apr 27, 2026
Use of uninitialized resources in vLLM before 0.19.1 allows remote attackers to corrupt inference output through concurrent API requests. The KV cache zeroing decision covers Mamba models but leaves recycled FullAttention block tails uncleared, allowing stale K/V values to propagate through masked softmax. A stock server without prefix caching is affected; four or five concurrent requests can yield nondeterministic output, with no demonstrated code execution or service interruption.
Summary
What happened
Use of uninitialized resources in vLLM before 0.19.1 allows remote attackers to corrupt inference output through concurrent API requests. The KV cache zeroing decision covers Mamba models but leaves recycled FullAttention block tails uncleared, allowing stale K/V values to propagate through masked softmax. A stock server without prefix caching is affected; four or five concurrent requests can yield nondeterministic output, with no demonstrated code execution or service interruption.
The record
- CVE
- CVE-2026-7141
- Published
- Apr 27, 2026
- Updated
- Apr 27, 2026
- Vendor
- vLLM Project
- Product
- vLLM
- Classifications
- CWE-908
- Attack vector
- network
- Privileges
- unauthenticated
Timeline
How it unfolded
- Apr 27, 2026CVE publishedPublication date reported by the CVE source.
- Apr 27, 2026Record updatedLatest update available in the CVE record.
Exploitability
Present is not the same as exploitable
Compare your product and version with the public record. A matching version still requires validation against your environment.
Is a vulnerable build present?
Compare these published version ranges with your installed build and any vendor patches.
- Affected versionversion=0.1
- Affected versionversion=0.10
- Affected versionversion=0.11
- Affected versionversion=0.12
- Affected versionversion=0.13
- Affected versionversion=0.14
- Affected versionversion=0.15
- Affected versionversion=0.16
- Affected versionversion=0.17
- Affected versionversion=0.18
- Affected versionversion=0.19.0
- Affected versionversion=0.2
- Affected versionversion=0.3
- Affected versionversion=0.4
- Affected versionversion=0.5
- Affected versionversion=0.6
- Affected versionversion=0.7
- Affected versionversion=0.8
- Affected versionversion=0.9
What conditions does exploitation require?
What is affected?
Attacks
What attackers are doing with it
Daily unique IPs observed by Shadowserver honeypots for known exploited vulnerabilities (KEVs). Missing observations do not establish an absence of attacks.
Weakness, pattern, technique
Public exploit references
- Concurrent-request KV corruption reproducerfunctional · demonstrated
Labels summarize the accepted research assessment. They do not indicate a test against your environment.
Technologies
Your stack
See the directory against your own environment.
Your stack
Check the software in your environment
Book a demo to see how Hinoki identifies affected software and validates exploitability in your environment.
Book a demo