CVE-2026-22773: vLLM is vulnerable to DoS in Idefics3 vision models via image payload with ambiguous dimensions
vLLM 0.6.4 up to but not including 0.12.0 crashes its engine when a client sends a 1x1 pixel image to a model that uses the Idefics3 architecture, such as SmolVLM-Instruct. The unhandled tensor error terminates the EngineCore process and the server. Upgrade to vLLM 0.12.0.
Published Jan 13, 2026 · Updated Jun 8, 2026 · Analysis Nov 1, 2026
What an attacker can do
A 1x1 image with shape (1, 1, 3) in height-width-channel order is ambiguous. The Idefics3 image processor takes it for channel-first format with shape (3, H, W) and miscounts the image patches. In vllm/model_executor/models/idefics3.py, _process_image_input then calls image_features.split(num_patches.tolist()) with a sum of 17 for a tensor whose first dimension is 9. PyTorch raises RuntimeError: split_with_sizes expects split_sizes to sum exactly to 9, nothing catches it, and EngineCore dies.
Any user who can submit an image to a served Idefics3-architecture model can take the server down; the CVSS vector lists low privileges, meaning an API key or account. The advisory demonstrated it on HuggingFaceTB/SmolVLM-Instruct. It affects availability only. A multi-tenant endpoint or a public demo will go down for every user until it restarts, and an attacker can repeat the request.
Who is affected
The vllm pip package from 0.6.4 up to 0.12.0 (exclusive), serving a model with the Idefics3 architecture. Check the version:
pip show vllmThen see which models your servers load. Other architectures are not named in the advisory.
vllm
- Ecosystem
- PyPI
- Affected
- >= 0.6.4, < 0.12.0
- Fixed in
- 0.12.0
- Check
pip show vllm
How to fix it
Upgrade to vLLM 0.12.0 or later:
pip install --upgrade "vllm>=0.12.0"The fix is pull request 29881. The advisory also sketches a mitigation for those who cannot upgrade: validate that image height and width exceed a minimum size before processing, and catch the RuntimeError around the split so a bad image gives an error response instead of ending the engine. As an interim step, a gateway can reject images below a minimum size.
How to tell if you were hit
Look in the vLLM logs for RuntimeError: split_with_sizes expects split_sizes to sum exactly to followed by the EngineCore process exiting. Check access logs for multimodal requests with very small images just before the crash, and for repeated restarts of the server. Note the client identity or API key behind those requests.
Original advisory
From the GitHub Advisory Database (CC BY 4.0).
Summary
Users can crash the vLLM engine serving multimodal models that use the _Idefics3_ vision model implementation by sending a specially crafted 1x1 pixel image. This causes a tensor dimension mismatch that results in an unhandled runtime error, leading to complete server termination.
Details
The vulnerability is triggered when the image processor encounters a 1x1 pixel image with shape (1, 1, 3) in HWC (Height, Width, Channel) format. Due to the ambiguous dimensions, the processor incorrectly assumes the image is in CHW (Channel, Height, Width) format with shape (3, H, W). This misinterpretation causes an incorrect calculation of the number of image patches, resulting in a fatal tensor split operation failure.
Crash location: vllm/model_executor/models/idefics3.py line 672:
def _process_image_input(self, image_input: ImageInputs) -> torch.Tensor | list[torch.Tensor]:
# ...
num_patches = image_input["num_patches"]
return [e.flatten(0, 1) for e in image_features.split(num_patches.tolist())]The split() call fails because the computed num_patches value (17) does not match the actual tensor dimension (9):
RuntimeError: split_with_sizes expects split_sizes to sum exactly to 9
(input tensor's size at dimension 0), but got split_sizes=[17]This unhandled exception terminates the EngineCore process, crashing the server.
Affected Models
Any model using the Idefics3 architecture. The vulnerability was tested with HuggingFaceTB/SmolVLM-Instruct.
Impact
Denial of service by crashing the engine
Mitigation
Validating the input:
def _validate_image_dimensions(self, image_shape):
h, w = image_shape[:2] if len(image_shape) == 3 else image_shape
if h < MIN_IMAGE_SIZE or w < MIN_IMAGE_SIZE:
raise ValueError(f"Image dimensions too small: {h}x{w}")Managing the exception:
try:
return [e.flatten(0, 1) for e in image_features.split(num_patches.tolist())]
except RuntimeError as e:
logger.error(f"Image processing failed: {e}")
raise InvalidImageError("Failed to process image features") from eFixes
- https://github.com/vllm-project/vllm/pull/29881
Credited to oxcabe, Isotr0py, DarkLight1337.
Questions
- Does it affect all vision models in vLLM?
The advisory says any model using the Idefics3 architecture, and it was tested with
HuggingFaceTB/SmolVLM-Instruct. It does not name other architectures.- Do I need an account to trigger it?
The CVSS vector says low privileges, so you need access to submit requests to the API, such as a key. Anyone with that access can crash the server.
- Can it execute code or leak data?
No. The advisory reports denial of service only, from an uncaught error that ends the engine process.
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 entriesCVE-2026-115911
vLLM: DoS via unbounded chat_template Jinja2 rendering exhausts the renderer thread pool
medium6.5Oct 28, 2026CVE-2026-115912
vLLM: Public multimodal SHM handle forgery reaches worker-side `pickle.loads()` with insecure serialization disabled
high7.5Oct 28, 2026CVE-2026-115913
vLLM: Phi-4 multimodal audio encoder materializes and retains full attention matrices: a single max-length audio OOM-crashes the vLLM engine
medium6.5Oct 28, 2026CVE-2026-115914
vLLM DeepSeek-OCR image processing vulnerable to denial of service
medium6.5Oct 28, 2026CVE-2026-115915
vLLM: Untrusted EPS inputs can trigger denial of service or Ghostscript RCE
high8.1Oct 28, 2026CVE-2026-115916
vLLM: Out-of-vocabulary prompt token id passes input validation and triggers a CUDA device-side assert (engine crash / DoS)
medium6.5Oct 28, 2026
