NewApp Monitoring is liveSign up free
AgentbeamBeta
AI agent vulnerability · vLLM

CVE-2026-44222: vLLM Vulnerable to Remote DoS via Special-Token Placeholders

mediumCVSS 3.1 6.5GHSA-hpv8-x276-m59f

vLLM 0.6.1 up to but not including 0.20.0 can be knocked offline by one text-only request that spells out image or video placeholder tokens when serving a vision-language model. The code then indexes empty grids, raises an IndexError, and can kill the worker. Upgrade to vLLM 0.20.0.

Published May 5, 2026 · Updated Jul 17, 2026 · Analysis Oct 28, 2026

What an attacker can do

Vision-language models such as Qwen2.5-VL use placeholder tokens like <|vision_start|><|image_pad|><|vision_end|> in the prompt, and image or video data supplied separately. The rotary position code in vllm/model_executor/layers/rotary_embedding.py (get_input_positions_tensor and _vl_get_input_positions_tensor) counts placeholder tokens in the input, then indexes image_grid_thw and video_grid_thw to match. If a user types the placeholder text but sends no image or video, those grids are empty and the code indexes them without a bounds check. The result is an unhandled IndexError (list index out of range).

The reporter reproduced it on vLLM 0.10.0 with Qwen/Qwen2.5-VL-3B-Instruct. A plain chat completion with the text "what's in picture <|vision_start|><|image_pad|><|vision_end|>" returned HTTP 500. In some deployments the worker exits, and capacity stays reduced until someone restarts it.

Any deployment that serves a vision-language model and accepts raw user text through the OpenAI-compatible API is exposed, including servers behind a proxy. The CVSS vector lists low privileges, and in practice that means anyone who can send a request. The effect is availability only.

Who is affected

vLLM from 0.6.1 up to but not including 0.20.0, serving multimodal models whose position handling uses image_grid_thw or video_grid_thw, such as Qwen2.5-VL. Text-only models do not take this path. Check your version and what you serve:

pip show vllm

vllm

Ecosystem
PyPI
Affected
>= 0.6.1, < 0.20.0
Fixed in
0.20.0
Check
pip show vllm

How to fix it

Upgrade to vLLM 0.20.0 or later:

pip install --upgrade "vllm>=0.20.0"

The advisory points to the tracking issue 32656, which moves M-RoPE position computation to use token indices from mm_features instead of searching input tokens for placeholders. That issue was closed in April 2026. The advisory names no workaround. If you cannot upgrade, you could filter placeholder strings such as <|image_pad|> and <|video_pad|> out of user text at a gateway; this is my suggestion, not one from the sources.

How to tell if you were hit

Look in vLLM logs for IndexError: list index out of range raised from _vl_get_input_positions_tensor, and for HTTP 500 responses on /v1/chat/completions. Check for worker exits or restarts shortly after a request. In request logs, look for prompts that contain vision placeholder strings without any image or video content part. Repeated 500s from one client are the likely pattern of abuse.

Original advisory

From the GitHub Advisory Database (CC BY 4.0).

Summary

This report explains a Token Injection vulnerability in vLLM’s multimodal processing. Unauthenticated, text-only prompts that spell special tokens are interpreted as control. Image and video placeholder sequences supplied without matching data cause vLLM to index into empty grids during input-position computation, raising an unhandled IndexError and terminating the worker or degrading availability. Multimodal paths that rely on image_grid_thw/video_grid_thw are affected. Severity: High (remote DoS). Reproduced on vLLM 0.10.0 with Qwen2.5-VL.

Details

  • Affected component: multimodal input position computation.
  • File/functions (paths are indicative):
  • vllm/model_executor/layers/rotary_embedding.py
  • get_input_positions_tensor(...)
  • _vl_get_input_positions_tensor(...)
  • Failure mechanism:
  • The code counts detected vision tokens and then indexes video_grid_thw/image_grid_thw accordingly.
  • When user input carries placeholder tokens but no actual multimodal payload, these grids are empty. The code does not bounds-check before indexing.

Representative snippet (context):

# vllm/model_executor/layers/rotary_embedding.py
@classmethod
def _vl_get_input_positions_tensor(
    cls,
    input_tokens,
    hf_config,
    image_grid_thw,
    video_grid_thw,
    ...,
):
    # detect video tokens
    video_nums = (vision_tokens == video_token_id).sum()
    # later in processing
    t, h, w = (
        video_grid_thw[video_index][0],  # IndexError if no video data
        video_grid_thw[video_index][1],
        video_grid_thw[video_index][2],
    )

Abbreviated call path:

OpenAI API request
 → vllm.v1.engine.core: step/execute_model
 → vllm.v1.worker.gpu_model_runner: _update_states/execute_model
 → vllm.model_executor.layers.rotary_embedding: get_input_positions_tensor
 → _vl_get_input_positions_tensor
 → IndexError: list index out of range

PoC

Environment

  • vLLM: 0.10.0
  • Model: Qwen/Qwen2.5-VL-3B-Instruct
  • Launch server:
python -m vllm.entrypoints.openai.api_server \
  --model Qwen/Qwen2.5-VL-3B-Instruct \
  --port 8000

Request (text-only, no image/video data)

cat > request.json <<'JSON'
{
  "model": "Qwen/Qwen2.5-VL-3B-Instruct",
  "messages": [
    {
      "role": "user",
      "content": [
        { "type": "text",
          "text": "what's in picture <|vision_start|><|image_pad|><|vision_end|>" }
      ]
    }
  ]
}
JSON

curl -s http://127.0.0.1:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  --data @request.json

Observed result

  • HTTP 500; logs show IndexError: list index out of range from _vl_get_input_positions_tensor(...).
  • In some deployments, the worker exits and capacity remains reduced until manual restart.

Impact

  • Type: Token Injection leading to Remote Denial of Service (unauthenticated). A single request can trigger the fault.
  • Scope: Any vLLM deployment that serves VLMs and accepts raw user text via OpenAI-compatible endpoints (self-hosted or proxied/managed fronts).
  • Effect: Request → unhandled exception in position computation → worker termination / service unavailability.

Fixes

  • Changes associated with https://github.com/vllm-project/vllm/issues/32656

Credits

Pengyu Ding (Infra Security, Ant Group) Ziteng Xu (Infra Security, Ant Group)

Credited to wumingzhilian.

Questions

Is vLLM 0.19.x affected by CVE-2026-44222?

Yes, if it serves a vision-language model. The affected range is 0.6.1 up to but not including 0.20.0.

Does the attacker need to upload an image?

No. The trigger is a text-only prompt that contains the placeholder tokens with no matching image or video data.

Do text-only models need patching?

The flaw sits in the multimodal position computation, so servers running only text models do not reach it. Upgrade anyway if you can.

References

How this page is made: the data panel comes from the GitHub Advisory Database, CISA’s KEV catalog, the NVD and FIRST EPSS, and updates twice a day. The analysis is written by an AI model from the sources listed above and checked by automated rules before it is published. Check versions against the vendor's advisory before you act. Methodology

More vLLM vulnerabilities

All vLLM entries