CVE-2025-47277

vLLM (versions 0.6.5–0.8.4) is affected only when using the PyNcclPipe KV cache transfer integration with the V0 engine. The issue stems from the PyTorch TCPStore binding defaulting to ALL interfaces; a workaround constrained the store to a private interface, and as of version 0.8.5 vLLM now bind...

CVE-2025-29770

CVE-2025-29770 affects vLLM’s guided_decoding outlines backend. The vulnerability arises because outlines_logits_processors.py unconditionally uses outlines’ on-disk grammar cache, enabling a malicious user to send many short decoding requests with unique schemas and exhaust the filesystem, causi...

CVE-2025-24357

The CVE-2025-24357 issue centers on vLLM’s hf_model_weights_iterator (vllm/model_executor/weight_utils.py) which loads checkpoints via torch.load with weights_only defaulting to False. If malicious pickle data is unpickled, arbitrary code could execute on the host. This vulnerability is highlight...

CVE-2025-29783

CVE-2025-29783 affects vLLM when Mooncake is configured for KV distribution across distributed hosts. The root cause is unsafe deserialization via pickle in the mooncake_pipe path exposed over ZMQ/TCP on all network interfaces, allowing remote code execution on affected distributed hosts. Public ...

CVE-2025-25183

CVE-2025-25183 affects vLLM (prefix cache) where malicious inputs can trigger Python 3.12’s hash(None) behaving as a predictable constant, enabling hash collisions in the prefix cache. This may allow cache entries created from one prompt to be reused for another, causing unintended behavior in re...

CVE-2025-30165

CVE-2025-30165 : In multi-node vLLM deployments using the V0 engine, a secondary host opens a SUB socket and deserializes inbound data with Python’s unsafe pickle over ZeroMQ XPUB/SUB, enabling remote code execution. The issue affects V0 deployments with tensor parallelism across hosts; V1 is una...

CVE-2025-32444

vLLM (0.6.5–0.8.4) with mooncake integration is vulnerable to remote code execution due to pickle-based serialization over unsecured ZeroMQ sockets that were listening on all interfaces. This could be exploited remotely; non-mooncake deployments are not affected. The issue is fixed in vLLM 0.8.5....

CVE-2025-46560

CVE-2025-46560 affects vLLM 0.8.0–0.8.4, where the multimodal tokenizer’s input preprocessing uses placeholder tokens replaced by repeated tokens. The replacement logic relies on inefficient list concatenation, yielding quadratic time complexity (O(n²)) and enabling resource exhaustion via crafte...

CVE-2025-30202

CVE-2025-30202 affects vLLM versions 0.5.2 up to 0.8.4 (prior to 0.8.5) in multi-node deployments. The root cause is an XPUB ZeroMQ socket bound to ALL interfaces on the primary host used for tensor parallelism, which can be accessed by any client with network access. This allows potential data e...

CVE-2026-48746

vLLM OpenAI auth bypass (CVE-2026-48746) affects vLLM versions 0.3.0 through 0.21.0. Root cause: ASGI servers and Starlette trust the Host header from the request scope, enabling manipulation of the reconstructed URL path and bypassing the OpenAI API AuthenticationMiddleware for routes beginning ...

CVE-2025-48943

The CVE-2025-48943 issue affects vLLM (inference/serving engine). Versions 0.8.0 up to but excluding 0.9.0 are vulnerable to a Denial of Service caused by processing an invalid regex when using structured output; the issue is fixed in 0.9.0. Related Chainguard advisories list affected builds such...

CVE-2025-46570

The CVE-2025-46570 entry concerns vLLM (inference/serving engine). The concrete detail across connected records shows a vulnerability in the PageAttention-based prefill path: when a new prompt is processed, a matching prefix chunk can accelerate prefill, creating timing differences (TTFT) that co...

CVE-2025-46722

The CVE-2025-46722 entry concerns vLLM (versions 0.7.0–0.8.x) where MultiModalHasher in vllm/multimodal/hasher.py hashes PIL.Image.Image objects using only obj.tobytes(). This excludes image metadata (width, height, mode), enabling two images with identical pixel data but different shapes to yiel...

CVE-2025-48887

CVE-2025-48887 affects vLLM command/tool parsing: the ReDoS vulnerability is in vllm/entrypoints/openai/tool_parsers/pythonic_tool_parser.py for versions 0.6.4 through 0.9.0 (exclusive) . The root cause is a highly complex, nested regex used for tool call detection, enabling catastrophic backtrac...

CVE-2025-48942

vLLM is an inference/serving engine for LLMs. CVE-2025-48942 affects versions 0.8.0 up to but excluding 0.9.0; hitting /v1/completions with an invalid json_schema as a Guided Param kills the vLLM server (availability impact). The issue is noted as analogous to CVE-2025-48943 but for a regex inste...

CVE-2024-11041

vllm-project vLLM version 0.6.2 contains a vulnerability in MessageQueue.dequeue() where pickle.loads is used to parse received sockets, enabling remote code execution if a malicious payload is sent to the MessageQueue. Multiple sources (CVE-2024-11041 entries across OSV, RH Red Hat, GHSA, CHAING...

CVE-2025-48944

vLLM (inference/serving engine) is affected when running versions 0.8.0 up to but excluding 0.9.0 with the /v1/chat/completions OpenAPI endpoint. The root cause is lack of validation for unexpected or malformed inputs in the pattern and type fields when the tools functionality is invoked, allowin...

CVE-2026-22778

Summary of CVE-2026-22778 : A vulnerability in vLLM (0.8.3–0.14.0) lets an attacker send an invalid image to the multimodal endpoint, causing PIL to leak a heap address. This information disclosure can be chained with a heap overflow in the JPEG2000 decoder used by OpenCV/FFmpeg to achieve remote...

CVE-2026-41523

vLLM prior to 0.22.0 is affected by an assert-based security check in the activation function loading that can permit arbitrary code execution when a malicious HuggingFace model is loaded and vLLM runs in Python optimized mode. The attacker-controlled inputs are the activation function names from...

CVE-2026-54236

CVE-2026-54236 affects vLLM versions before 0.23.1rc0. Five code paths bypass the sanitize_message global exception handler, leaking heap addresses via exception messages: (1) Anthropic API router POST /v1/messages and POST /v1/messages/count_tokens (vllm/entrypoints/anthropic/api_router.py), (2)...

CVE-2026-5497

CVE-2026-5497 affects vLLM 0.8.0 and later, where VideoMediaIO.load_base64() can perform unbounded frame processing for video/jpeg data URLs, leading to an Out-of-Memory DoS. An attacker can craft a single API request with thousands of comma-separated base64 JPEG frames, causing the server to dec...

CVE-2026-54232

vLLM prior to 0.22.1 is affected by a dependency confusion flaw in its Dockerfile. The vulnerability arises from installing flashinfer-jit-cache from a private index (flashinfer.ai/whl/) via --extra-index-url while the package name was not registered on PyPI and UV_INDEX_STRATEGY is set to unsafe...

CVE-2026-55574

CVE-2026-55574 affects vLLM prior to 0.24.0, where structured_outputs.regex passes an unguarded user-supplied regex to grammar backends (xgrammar and outlines). In xgrammar, the string reaches the regex compiler without a timeout guard; in outlines, validation overlooks regex complexity (e.g., ne...

CVE-2025-48956

Technical details for CVE-2025-48956 are not publicly available in the provided documents. Monitor for updates from project advisories; no verified affected versions, exploit status, or remediation details are included here.

CVE-2026-55646

CVE-2026-55646 (vLLM) affects vLLM versions 0.22.0–0.23.0. The routes /v1/audio/transcriptions and /v1/audio/translations call request.file.read() to fully materialize an uploaded audio file before enforcing the documented size limit (default 25 MB). This can cause the server to allocate memory p...

CVE-2026-55514

CVE-2026-55514 affects the vLLM library (inference/serving) from versions 0.12.0 through older than 0.24.0. Sending a pure prompt embeds payload in a /v1/completions request for a model using M-RoPE triggers an EngineCore assertion, causing a fatal crash that shuts down the entire server applicat...

CVE-2026-54234

CVE-2026-54234 affects vLLM prior to version 0.24.0. A frontend multi-request speculative decoding path could cause the rejection sampler to produce a recovered token equal to the model vocabulary size boundary value, which becomes negative one when the next live token is selected. This out-of-vo...

CVE-2026-47155

CVE-2026-47155 affects vLLM prior to 0.22.0. Description: revision pinning controls do not consistently apply to all artifacts loaded for a model, enabling loading of dynamic code, GGUF files, image processors, retrieval side weights, or same-repository subfolder weights/config from an unpinned/d...

CVE-2026-25960

Summary of CVE-2026-25960 (vLLM) : The SSRF protection added in 0.15.1 (fix tied to CVE-2026-24779) can be bypassed in vLLM’s load_from_url_async due to inconsistent URL parsing between the validation layer (urllib3.util.parse_url) and the HTTP client (aiohttp with yarl). The vulnerability arises...

CVE-2026-27893

CVE-2026-27893 affects vLLM’s inference/serving engine. From version 0.10.1 up to (but not including) 0.18.0, two model implementation files hardcode trust_remote_code=True when loading sub-components, bypassing the user’s --trust-remote-code=False security opt-out. This enables remote code execu...

CVE-2026-34756

CVE-2026-34756 affects vLLM OpenAI-compatible API server prior to 0.19.0. The root cause is missing upper-bound validation on the n parameter in ChatCompletionRequest/CompletionRequest, allowing an unauthenticated attacker to send an astronomically large n value that causes the asyncio event loop...

CVE-2026-54235

Summary: CVE-2026-54235 affects vLLM prior to 0.23.1rc0, where temperature validation gates using can silently mis-handle NaN and positive Infinity due to Python IEEE 754 behavior. This allows non-finite temperatures to bypass guards and propagate to GPU sampling kernels, causing undefined behav...

CVE-2025-62164

The CVE affects vLLM (inference/serving engine) before 0.11.1, where the Completions API loads user-supplied prompt embeddings with torch.load() lacking proper validation. A PyTorch 2.8.0 change disables sparse-tensor invariants checks, allowing crafted tensors to bypass bounds checks and trigger...

CVE-2026-44223

vLLM contains a vulnerability (CVE-2026-44223) where the extract_hidden_states speculative decoding pathway can crash the EngineCore process if any request uses penalty parameters (repetition_penalty, frequency_penalty, or presence_penalty). The issue arises from an incorrect tensor shape after t...

CVE-2026-56340

vLLM versions >= 0.10.2 and

CVE-2026-54233

Affected software: vLLM (inference/serving engine). Vulnerability: decoding an audio file on the /v1/audio/transcriptions endpoint can cause extreme memory growth. A 25 MB OPUS upload decodes to about 14.9 GB of float32 PCM, because the audio decoder concatenates all frames in memory before retur...

CVE-2026-24779

CVE-2026-24779 is an SSRF vulnerability in vLLM’s MediaConnector. Before version 0.14.1, load_from_url and load_from_url_async fetch media from user-supplied URLs and validate via Python urllib urlparse, while the request is issued with requests/urllib3, whose parsing follows a different standard...

CVE-2026-22807

Vulnerability CVE-2026-22807 affects vLLM versions prior to 0.14.0, where during model resolution the engine loads Hugging Face auto_map dynamic modules without gating on trust_remote_code. This allows attacker-controlled Python code in a model repo or path to execute at server startup, before an...

CVE-2025-66448

vLLM (prior to 0.11.1) contains a remote code execution vulnerability in Nemotron_Nano_VL_Config where, during model loading, an auto_map entry can cause get_class_from_dynamic_module to fetch and execute code from a remote repository, bypassing trust_remote_code checks. This can enable an attack...

CVE-2026-34755

vLLM's VideoMediaIO.load_base64("video/jpeg") path has an unbounded frame-splitting bug: data.split(",") bypasses the intended frame-count limit (default 32) used by the binary path, allowing a single request with thousands of comma-separated base64 JPEG frames. This can cause the server to decod...

CVE-2025-59425

CVE-2025-59425 affects vLLM prior to 0.11.0rc2, where API key validation used a string comparison that leaks timing information. Attackers observing responses over many attempts could determine the next correct character in the API key, enabling authentication bypass. The issue is resolved in ver...

CVE-2026-34760

Summary: CVE-2026-34760 concerns vLLM’s audio processing path via Librosa. From version 0.5.5 up to before 0.18.0, Librosa used numpy.mean for mono downmix (to_mono), while ITU-R BS.775-4 specifies a weighted downmix. This mismatch creates inconsistency between audio perceived by humans and audio...

CVE-2026-53923

Summary of CVE-2026-53923 : The vulnerability affects vLLM (GGUF dequantize kernels) where integer truncation of tensor dimensions causes partially filled output tensors. From 0.5.5 up to 0.23.1rc0, the code allocates the full output tensor (torch::empty) but the CUDA kernel processes only a trun...

CVE-2025-62372

CVE-2025-62372 affects vLLM (inference/serving engine). From version 0.5.5 up to before 0.11.1, passing multimodal embedding inputs with correct ndim but incorrect shape (e.g., wrong hidden dimension) can crash the engine when serving multimodal models, regardless of whether those inputs are supp...

CVE-2026-44222

CVE-2026-44222 (vLLM) affects vLLM versions 0.6.1 through 0.19.x where a token-injection vulnerability in multimodal processing allows unauthenticated text prompts containing special tokens to be interpreted as control. When image/video placeholder sequences are provided without corresponding dat...

CVE-2026-22773

CVE-2026-22773 affects vLLM (inference/serving engine) versions 0.6.4 through before 0.12.0 that serve multimodal models using the Idefics3 vision model. A crafted 1x1 pixel image triggers a tensor dimension mismatch in the image input processing, causing an unhandled runtime error and enabling a...

CVE-2026-34753

The CVE-2026-34753 issue is an SSRF in vLLM’s download_bytes_from_url used by the batch runner. Attacker-controlled batch input JSON can set file_url to arbitrary HTTP(S) URLs, and the code fetches them without URL validation, allowing requests from the vLLM host to internal or external targets (...

CVE-2025-62426

Summary: CVE-2025-62426 affects vLLM up to versions before 0.11.1. The /v1/chat/completions and /tokenize endpoints accept a chat_template_kwargs parameter that is used before validation, allowing an attacker to block the API server by forcing large tokenization tasks and delaying all other reque...

CVE-2026-7141

CVE-2026-7141 affects vllm up to 0.19.0, specifically the KV Block Handler’s has_mamba_layers function in vllm/v1/kv_cache_interface.py. A manipulation can trigger an uninitialized resource, with remote initiation possible. The issue is described as high complexity and difficult exploitability, w...

CVE-2025-71379

CVE-2025-71379 affects vLLM versions 0.6.3 through 0.8.x (before 0.9.0). The vulnerability is a set of regular expression denial of service (ReDoS) flaws in multiple components: (1) regex patterns in vllm/lora/utils.py, (2) the phi4mini tool parser, and (3) the OpenAI-compatible serving chat endp...