Cross-user inference disclosure via CUDA offset overflow
Published Aug 13, 2026 · Updated Aug 13, 2026
Integer overflow in vLLM before 0.27.0 allows remote attackers to read another user's inference result through a shared batch. In act_and_mul_kernel, 32-bit evaluation of blockIdx.x * 2 * d wraps before pointer addition, so the CUDA kernel reads a different batch element's input. Exploitation requires attacker and victim requests in the same inference batch with dimensions that trigger the wrap, exposing part or all of the victim's generated output.
Summary
What happened
Integer overflow in vLLM before 0.27.0 allows remote attackers to read another user's inference result through a shared batch. In act_and_mul_kernel, 32-bit evaluation of blockIdx.x * 2 * d wraps before pointer addition, so the CUDA kernel reads a different batch element's input. Exploitation requires attacker and victim requests in the same inference batch with dimensions that trigger the wrap, exposing part or all of the victim's generated output.
The record
- CVE
- CVE-2026-73558
- Published
- Aug 13, 2026
- Updated
- Aug 13, 2026
- Vendor
- vLLM Project
- Product
- vLLM
- Classifications
- CWE-190, T1190
- Attack vector
- network
- Privileges
- unauthenticated
Timeline
How it unfolded
- Aug 13, 2026CVE publishedPublication date reported by the CVE source.
- Aug 13, 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=< 0.27.0
What conditions does exploitation require?
What is affected?
Published CVSS scores
CVSS describes severity. EPSS estimates exploitation probability.
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
- Upstream shared-batch reproductionproof of concept · 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