CVE-2026-100649: vLLM before 0.29.0 Resource Limit Bypass via Sampler Subclass
vLLM before 0.29.0 contains a resource-limit bypass vulnerability in PyNvVideoCodec decoder allocation where sampler subclass shadowing allows independent counter increments. Unauthenticated attackers can select different sampler subclasses in video requests to exceed configured decoder limits and exhaust unaccounted GPU memory.
Affected Software
Remediation
Recommended actions to resolve this vulnerability, in priority order.
- Upgrade
Upgrade
vLLMto a version that resolves this vulnerability.Fixed in 0.29.0
Event History
Frequently Asked Questions
Which deployments should be prioritized for remediation?
Deployments running vLLM versions before 0.29.0 that accept video requests are affected. Exposure is greatest where unauthenticated users can submit such requests.
Does exploitation require credentials or user interaction?
No. The issue can be exploited by unauthenticated attackers, and no user interaction is required.
What is the practical impact of successful exploitation?
An attacker can use different sampler subclasses to bypass configured decoder limits, causing GPU memory allocations that are not accounted for. This can exhaust GPU memory and affect availability.
How can I determine whether my deployment is vulnerable?
Check the installed vLLM version. Versions before 0.29.0 are affected; version 0.29.0 and later are not identified as affected by the provided advisory information.