vLLM is an inference and serving engine for large language models (LLMs). In versions from 0.6.4 to before 0.12.0, 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. This issue has been patched in version 0.12.0.

Affected: PyPI: vllm

### 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: ```python 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 `Hug…

Affected: PyPI: vllm

vLLM is an inference and serving engine for large language models (LLMs). In versions from 0.6.4 to before 0.12.0, 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. This issue has been patched in version 0.12.0.

vllm: Server Side request forgery (SSRF) in MediaConnector feast: Feast: Remote Code Execution via insecure YAML deserialization openshift-ai: Trusty AI Grants All Authenticated users to list pods in any namespace nltk: Zip Slip Vulnerability in nltk Leading to Code Execution golang: net/url: Memory exhaustion in query parameter parsing in net/url golang: archive/zip: Excessive CPU consumption when building archive index in archive/zip crypto/x509: golang: Denial of Service due to excessive resource consumption via crafted certificate urllib3: urllib3: Unbounded decompression chain leads to resource exhaustion urllib3: urllib3 Streaming API improperly handles highly compressed data cbor2: cbor2: Information Disclosure via shared memory in CBORDecoder reuse aiohttp: AIOHTTP's HTTP Parser auto_decompress feature is vulnerable to zip bomb aiohttp: aiohttp: Denial of Service via specially crafted POST request aiohttp: aiohttp: Denial of Service via memory exhaustion from crafted POST request python-markdown: denial of service via malformed HTML-like sequences nltk: NLTK: Arbitrary file read via improper path validation in `filestring()` function nltk: NLTK: Arbitrary file read via path traversal vulnerability io.vertx/vertx-core: static handler component cache can be manipulated to deny the access to static files google-cloud-aiplatform: google-cloud-aiplatform: Arbitrary code execution via Stored Cross-Site Scripting (XSS) tensorflow: TensorFlow: Local privilege escalation via…

Ein entfernter, authentisierter Angreifer kann eine Schwachstelle in vllm ausnutzen, um einen Denial of Service Angriff durchzuführen.