CVE-2026-100649: vLLM: sampler subclass bypass exhausts GPU memory
MEDIUMvLLM before 0.29.0 has a resource-limit bypass in its PyNvVideoCodec decoder allocation: because of sampler subclass shadowing, each subclass increments its own counter, so an unauthenticated remote client can request video decoding through different sampler subclasses and exceed the configured decoder limits. The result is unaccounted GPU memory exhaustion and degraded or interrupted inference, with no impact on confidentiality or integrity. Risk is moderate: CVSS 5.3 (availability only), EPSS 0.00341 (top 75th percentile), not in CISA KEV, no public exploit or Nuclei template, and CISA SSVC says TRACK. The exposure that matters is vLLM servers that accept video input and are reachable by untrusted clients, where one request pattern can take down a shared GPU. Upgrade to vLLM 0.29.0 or later at the next maintenance window, and in the meantime restrict access to video-capable endpoints and enforce per-client rate limits.
What is the risk?
Medium. The vulnerability is network-exploitable with low complexity and needs no privileges or user interaction (AV:N/AC:L/PR:N/UI:N), but it only affects availability (A:L). Exploitation likelihood is low to moderate: EPSS is 0.00341, there is no public exploit, no scanner template and no KEV listing, and SSVC is TRACK. Exposure is highest for internet-facing or multi-tenant vLLM deployments with video/multimodal input enabled, where GPU memory is a shared and costly resource. Deployments without video decoding via PyNvVideoCodec are not exposed. Downstream dependents and package risk data are unknown.
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 0.29.0 or later (see GHSA-j682-9xp5-rrf3 and the VulnCheck advisory). 2) Until patched, disable video input or the PyNvVideoCodec decode path if it is not needed. 3) Put video-capable endpoints behind authentication, an API gateway or a reverse proxy with per-client rate limits and concurrency caps, and limit request body size. 4) Isolate vLLM on dedicated GPUs, and keep GPU memory headroom and container or cgroup limits. 5) Detection: monitor GPU memory utilisation (nvidia-smi/DCGM) against the configured decoder limits, alert on OOM errors and on bursts of video requests from a single source, and review request logs for requests that vary sampler parameters or types.
What does CISA's SSVC say?
Source: CISA Vulnrichment (SSVC v2.0). Decision based on the CISA Coordinator decision tree.
How is it classified?
Which compliance frameworks are affected?
This CVE is relevant to:
Frequently Asked Questions
What is CVE-2026-100649?
vLLM before 0.29.0 has a resource-limit bypass in its PyNvVideoCodec decoder allocation: because of sampler subclass shadowing, each subclass increments its own counter, so an unauthenticated remote client can request video decoding through different sampler subclasses and exceed the configured decoder limits. The result is unaccounted GPU memory exhaustion and degraded or interrupted inference, with no impact on confidentiality or integrity. Risk is moderate: CVSS 5.3 (availability only), EPSS 0.00341 (top 75th percentile), not in CISA KEV, no public exploit or Nuclei template, and CISA SSVC says TRACK. The exposure that matters is vLLM servers that accept video input and are reachable by untrusted clients, where one request pattern can take down a shared GPU. Upgrade to vLLM 0.29.0 or later at the next maintenance window, and in the meantime restrict access to video-capable endpoints and enforce per-client rate limits.
Is CVE-2026-100649 actively exploited?
No confirmed active exploitation of CVE-2026-100649 has been reported, but organizations should still patch proactively.
How to fix CVE-2026-100649?
1) Upgrade vLLM to 0.29.0 or later (see GHSA-j682-9xp5-rrf3 and the VulnCheck advisory). 2) Until patched, disable video input or the PyNvVideoCodec decode path if it is not needed. 3) Put video-capable endpoints behind authentication, an API gateway or a reverse proxy with per-client rate limits and concurrency caps, and limit request body size. 4) Isolate vLLM on dedicated GPUs, and keep GPU memory headroom and container or cgroup limits. 5) Detection: monitor GPU memory utilisation (nvidia-smi/DCGM) against the configured decoder limits, alert on OOM errors and on bursts of video requests from a single source, and review request logs for requests that vary sampler parameters or types.
What systems are affected by CVE-2026-100649?
This vulnerability affects the following AI/ML architecture patterns: model serving, LLM inference servers, multimodal/video inference pipelines, shared GPU clusters.
What is the CVSS score for CVE-2026-100649?
CVE-2026-100649 has a CVSS v3.1 base score of 5.3 (MEDIUM). The EPSS exploitation probability is 0.34%.
What is the AI security impact?
Affected AI Architectures
MITRE ATLAS Techniques
AML.T0029Denial of AI ServiceAML.T0034.001Resource-Intensive QueriesAML.T0040AI Model Inference API AccessAML.T0049Exploit Public-Facing ApplicationCompliance Controls Affected
What are the technical details?
Original Advisory
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.
Exploitation Scenario
An attacker finds a publicly exposed vLLM OpenAI-compatible endpoint that serves a video-language model. With no authentication needed, they send a stream of video requests, each selecting a different sampler subclass, so every subclass increments its own decoder counter and none reaches the configured limit. Decoder instances pile up and consume GPU memory outside vLLM's accounting until the GPU runs out of memory. Legitimate inference requests then fail or time out, and the service needs a restart, which the attacker can repeat at will.
Weaknesses (CWE)
CWE-770Allocation of Resources Without Limits or ThrottlingPrimaryCWE-770Allocation of Resources Without Limits or ThrottlingCWE-770 — Allocation of Resources Without Limits or Throttling: The product allocates a reusable resource or group of resources on behalf of an actor without imposing any intended restrictions on the size or number of resources that can be allocated.
- [Requirements] Clearly specify the minimum and maximum expectations for capabilities, and dictate which behaviors are acceptable when resource allocation reaches limits.
- [Architecture and Design] Limit the amount of resources that are accessible to unprivileged users. Set per-user limits for resources. Allow the system administrator to define these limits. Be careful to avoid CWE-410.
Source: MITRE CWE corpus.
CVSS Vector
CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:LReferences
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