vLLM vulnerabilities

vLLM before 0.29.0 passes uploaded audio to the PyAV/FFmpeg av.open() fallback with only a file-size check, exposing FFmpeg's IAMF parser heap overflow. Authenticated remote attackers can repeatedly upload malformed IAMF containers to /v1/audio/transcriptions, corrupting the heap and crashing the API process.

Vulnerability in vLLM versions 0.3.0 to 0.22.0 allows bypass of OpenAI API AuthenticationMiddleware. An attacker can use the API without providing the configured VLLM_API_KEY or --api-key.

A vulnerability in vLLM from version 0.8.3 to 0.14.0 leaks a heap address when an invalid image is sent to the multimodal endpoint. This leak reduces ASLR effectiveness from 4 billion to about 8 guesses, facilitating further attacks.

A security flaw has been discovered in vllm-project vLLM up to 0.31.0, affecting the function get_token_bin_counts_and_mask in the file vllm/model_executor/layers/utils.py of the Penalty Handler component. Manipulation results in denial of service. Remote exploitation is possible, and the exploit has been publicly released.

A security vulnerability in vllm-project vLLM up to 0.31.0 causes an out-of-bounds read in the function conv_ssm_forward in the file mamba_mixer2.py, within the Completions Request Handler component. The attack can be carried out remotely, and the exploit has been publicly disclosed.

In vLLM before 0.30.0, a caller can use the media_io_kwargs field to select the GLMGA video backend and supply large fps and max_frames values without a strict work ceiling, causing disproportionate CPU and memory consumption.

In vLLM before 0.30.0, the Rust frontend's track_http_metrics middleware records the raw HTTP method token as a Prometheus label, allowing an unauthenticated attacker to send unique arbitrary method tokens and cause permanent creation of label sets, increasing memory usage.

In vLLM from 0.24.0 to 0.30.0, the Qwen2VLVideoBackend and Qwen3VLVideoBackend classes accept request-level values for max_frames and fps without server-side ceilings, allowing an unauthenticated caller to decode every frame from attacker-controlled video and consume disproportionate memory.

In vLLM before 0.30.0, structured-output request failures can escape validation and reach the EngineCore fatal-error path, allowing ordinary constrained-generation requests to terminate the shared engine.

In vLLM before 0.30.0, OpenAI-compatible request models accept a non-empty cache_salt value without enforcing character and length restrictions, which can raise an uncaught ValueError during scheduler cache lookup, causing EngineCore to terminate.

In vLLM before 0.30.0, flash late-interaction scoring at /score and /rerank derives query_key from the caller-controlled X-Request-Id header, allowing a concurrent request to overwrite the cached query embedding of a victim.

In vLLM before 0.30.0, the /inference/v1/generate endpoint in the disaggregated scale-out path accepts caller-supplied tensors, cache identifiers, ranges, and wire-selected multimodal field processors without rebinding them to the active model renderer contract, which can terminate EngineCore, poison cache, or alter transport semantics.

In vLLM before 0.28.0, the default mirrored multimodal LRU cache can commit a media hash in the frontend sender cache during rendering and before engine admission, while the receiver cache never receives the payload if the request is rejected, causing an assertion error on later reuse.

vLLM is an inference and serving engine for large language models. Prior to 0.30.0, Harmony tool continuations submitted through POST /v1/responses requests rebuild the next-turn engine input without preserving the cache_salt value, placing the continuation prefix in the global unsalted cache namespace even when the caller enabled salting. On deployments with prefix caching enabled (default), an authenticated tenant who can reconstruct a victim's low-entropy post-tool history can submit the same continuation and use the cached_tokens_per_turn count to determine whether the prefix was previously processed, defeating tenant isolation.

A flaw has been found in vllm-project vLLM up to 0.26.0. This vulnerability affects unknown code of the file rust/src/parser/src/unified/gemma4.rs of the component Gemma4UnifiedParser. Executing a manipulation can lead to denial of service. The attack may be launched remotely. The exploit has been published and may be used. Upgrading to version 0.29.1rc0 is able to resolve this issue. This patch is called 3439bad37e68ba9755a46f4f6b44a4aeaf1f60a9. Upgrading the affected component is advised.

vLLM before 0.29.0 accepts user-controlled stop_token_ids on the OpenAI-compatible POST /v1/completions and POST /v1/chat/completions endpoints but validates only that the values are integers, not that each token id is within the model vocabulary/logits range. When min_tokens > 0, the stop token ids are used as logits indices to suppress stop tokens, so an out-of-range id reaches a CUDA indexing operation (index_put_) and triggers a device-side assertion. An authenticated API user can send a single malformed completion request that returns 500 Internal Server Error and puts EngineCore into a fatal state, causing subsequent requests to fail until the service is restarted (denial of service).

In vLLM versions 0.22.1 through 0.28.0, the operator-supplied model revision pin (--revision / --code-revision) is not propagated to several Hugging Face artifact loads for the FunAudioChat and Tarsier2 architectures. This causes pinned deployments to still resolve behavior-affecting artifacts from the repository's default revision, potentially altering audio preprocessing, speech tokenizer behavior, or Tarsier2 configuration without any change to the operator's configured pin. This is a supply-chain integrity and reproducibility failure, but not remote code execution or a trust_remote_code=False bypass. Fixed in version 0.28.0.

vLLM versions 0.22.0 through 0.23.0 fail to validate stop_token_ids against vocabulary bounds in Rust HTTP and gRPC frontends, allowing out-of-vocabulary token IDs to reach MinTokensLogitsProcessor. Attackers can submit requests with min_tokens greater than zero and out-of-vocabulary stop_token_ids to trigger CUDA tensor indexing failures that leave EngineCore in a fatal state requiring service restart.

vLLM before 0.29.0 fails to enforce decoder prompt-length validation on the disaggregated serving endpoint /inference/v1/generate. When the request contains a 'features' (multimodal) payload, caller-supplied token_ids are not checked against model_config.max_model_len, and for multimodal processors reporting skip_prompt_length_check=True an overlong prompt reaches the worker input-batch copy into a fixed max_model_len-wide NumPy row. A client able to reach the endpoint can trigger a worker failure and denial of service.

vLLM through 0.29.0 fetches and fully materializes remote or inline media before enforcing its documented media controls (the VLLM_MAX_AUDIO_CLIP_FILESIZE_MB compressed-audio size cap, default 25 MB, and per-modality --limit-mm-per-prompt item limits). Across four ingress paths the server reads the entire HTTP response body, base64-decodes the inline payload, or spawns one fetch/decode task per media part, and only then applies the limit (or never applies it). A remote attacker can cause pre-inference memory and bandwidth exhaustion (denial of service).

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.

vLLM before 0.29.0 fails to enforce the VLLM_MAX_AUDIO_CLIP_FILESIZE_MB limit in multimodal chat audio decoding, allowing unauthenticated clients to bypass file size restrictions. Attackers can submit oversized audio files through chat endpoints to consume excessive memory and CPU resources during decoding.

vLLM versions before 0.29.0 contain a denial-of-service vulnerability in the cache_salt parameter accepted on OpenAI-compatible and Anthropic API endpoints, which lacks maximum length validation and is processed on the single EngineCore scheduler thread. Unauthenticated attackers can send HTTP requests with multi-hundred-megabyte salt values that trigger expensive pickle serialization and SHA-256 hashing, stalling the scheduler thread and denying service to all concurrent requests.

vLLM through 0.29.0 fails to validate the tp_size parameter in kv_transfer_params on OpenAI-compatible completion endpoints, allowing attackers to allocate unbounded memory. Attackers can supply arbitrary tp_size values in prefill/decode disaggregated deployments to exhaust memory and trigger kernel OOM-kill of the decode worker process.

vLLM through 0.29.0 contains a resource exhaustion vulnerability in MooncakeConnector where rejected prefill requests create ownerless transfer placeholders that are never reclaimed. Attackers can send rejected requests to exhaust sender task pools, causing valid requests to be delayed by up to 480 seconds while health checks continue returning success.

vLLM through 0.29.0 contains a denial of service vulnerability in P2P KV offloading when OffloadingConnector is configured with TieringOffloadingSpec and a peer-to-peer secondary tier. Attackers can supply arbitrary remote host and port values in kv_transfer_params to create unreachable peer sessions that retain ZeroMQ sockets until the context quota is exhausted, causing an uncaught ZMQError that crashes EngineCore and stops all inference.

vLLM through 0.29.0 contains a denial of service vulnerability in the NIXL connector's prefix caching implementation that fails to properly validate block counts across multi-prompt completion requests in prefill/decode disaggregated deployments. Attackers can trigger an assertion failure in NixlBaseConnectorWorker._apply_prefix_caching by submitting completion requests with multiple prompts of varying lengths, causing the decode worker to terminate and become unavailable until restarted.

vLLM versions through 0.29.0 contain a denial of service vulnerability in the NIXL connector's metadata handling for prefill/decode disaggregated deployments. Attackers can send requests with incomplete kv_transfer_params dictionary entries to trigger an uncaught KeyError in EngineCore scheduling, causing the decode engine to terminate and making all routed requests fail until manual restart.

vLLM through 0.29.0 fails to properly validate bad_words token indices against the model's generation output width in SamplingParams.update_from_tokenizer(). Attackers can supply out-of-bounds token indices that corrupt logits memory of concurrent requests.

vLLM through 0.29.0 contains a memory corruption vulnerability in the Triton _bincount_kernel where prompt token IDs index the penalty prompt-presence bitset without bounds checking against vocabulary size. Attackers can submit multimodal audio requests with tokens equal to vocabulary size, causing out-of-bounds writes that corrupt concurrent requests' sampler state and alter repetition penalty behavior.

vLLM before 0.29.0 validates allowed_token_ids against tokenizer length instead of model output logits width in SamplingParams._validate_allowed_token_ids(). Attackers can supply token IDs above the output vocabulary that pass validation, causing LogitBiasState to corrupt GPU logits state and allow concurrent requests to sample tokens outside their allowlists.

vLLM versions before 0.28.0 fail to validate the lower bound of token IDs in the /v1/embeddings and /pooling endpoints, allowing unauthenticated attackers to crash the engine by submitting negative token IDs. A single request with a negative token ID triggers a CUDA device-side assertion that poisons the GPU context, causing all subsequent requests to fail until the process restarts.

vLLM through 0.29.0 fails to properly clean up decode-side metadata for rejected inference requests in prefill/decode disaggregated deployments. Remote attackers can submit requests with max_tokens=0 to exhaust decode-worker memory without bound until the worker restarts.

Prior to 0.28.0, request bodies for Chat Completions and Responses can set media_io_kwargs.video.video_backend to pynvvideocodec, and MediaConnector.fetch_video forwards that choice to VideoMediaIO even when startup configuration selected a software decoder. The engine's _reserve_mm_ipc_gpu_memory logic budgets decoder memory only from static configuration, so the request-selected VIDEO_LOADER_REGISTRY backend can create a CUDA context, decoder surfaces, and decoded-frame allocations that were not removed from the engine's KV-cache budget.

vLLM before version 0.24.0 in the input_audio handling path for /v1/chat/completions does not pass VLLM_MAX_AUDIO_DECODE_DURATION_S to the shared audio decoder. An unauthenticated client can submit a small compressed audio input that expands into a very large float32 PCM allocation, bypassing the duration guard and causing an out-of-memory worker crash.

A vulnerability was found in vllm-project vllm up to 0.29.0. Affected by this issue is some unknown functionality of the file vllm/v1/sample/thinking_budget_state.py. The manipulation results in inefficient algorithmic complexity, and the attack can be launched remotely.

A vulnerability was found in vLLM versions 0.26.0 and 0.27.0 within the MoRIIO Acknowledgement Handler component. Manipulation of the request_id/kv_transfer_params arguments in MoRIIOConnectorScheduler.request_finished, MoRIIOConnectorWorker.get_finished, and MoRIIOWrapper._handle_release_message leads to resource consumption. The attack can be initiated remotely, and the project has not yet reacted to the report.

In vLLM up to 0.29.0, a vulnerability in the TiktokenTokenizer::new function (rust/src/text/src/backend/hf/mod.rs) leads to DoS via local manipulation.

vLLM versions before 0.28.0 fail to validate audio sample rate headers in the transcription endpoint, allowing authenticated clients to bypass duration checks. Attackers can submit forged FLAC headers with inflated sample rates to trigger excessive memory allocation and crash the API server process affecting all tenants.

vLLM versions >=0.10.2 and <0.28.0 do not apply any audio decode-size or duration limit when extracting audio from video input for NanoNemotronVL models. In nano_nemotron_vl.py, _extract_audio_from_videos calls load_audio_pyav(BytesIO(video_bytes)) without the max_duration_s or max_decode_bytes parameters, so neither VLLM_MAX_AUDIO_DECODE_DURATION_S nor VLLM_MAX_AUDIO_DECODE_BYTES is enforced (unlike the direct audio upload path in AudioMediaIO). When a NanoNemotronVL model is served with use_audio_in_video=True, an attacker who supplies a small, highly compressed video as multimodal input can force the server to allocate gigabytes of memory during audio decoding, resulting in a denial of service. Fixed in vLLM 0.28.0.

vLLM before 0.28.0 contains a remote code execution vulnerability in the LlavaOnevision2 processor loader that ignores the trust_remote_code parameter when loading remote processor classes. Attackers can craft a malicious model with arbitrary code in processing_llava_onevision2.py that executes with vLLM process authority even when trust_remote_code is set to False.

vLLM up to and including 0.17.0 allows remote attackers to cause a Denial of Service via memory exhaustion. The AsyncMediaIO.fetch_audio and AsyncMediaIO.fetch_image functions in multimodal/inputs.py fetch user-supplied media URLs using aiohttp and call r.read() without enforcing a maximum response size, allowing an attacker to exhaust server memory by providing a URL to an arbitrarily large file.

vLLM before 0.27.0 fails to properly classify DeepStream as a GPU backend and omits pixel-limit enforcement in its decode path. Unauthenticated attackers can activate DeepStream at request time to initialize the process-wide GPU decode pool and submit video that bypasses resource controls, causing partial denial of service for concurrent requests.

vLLM prior to 0.26.0 in the MiMoV2OmniMultiModalProcessor passes attacker-controlled image and audio strings through _fetch_image, requests.get, and Image.open instead of MediaConnector, bypassing allowed_media_domains and allowed_local_media_path protections. This allows server-side requests and reads of arbitrary files accessible to the vLLM process.

vLLM is an inference and serving engine for large language models. Prior to 0.26.0, the /v1/completions/derender and /v1/chat/completions/derender endpoints accept caller-supplied GenerateResponse objects whose generate_responses, choices, token_ids, prompt_logprobs, logprobs.content, top_logprobs, and routed_experts structures are processed by OnlineDerenderer and tokenizer.decode before max_model_len, max_tokens, max_num_seqs, or response-size limits are enforced, allowing an authenticated API client to consume excessive CPU and memory and produce oversized responses. This issue is fixed in version 0.26.0.

vLLM is an inference and serving engine for large language models. From 0.19.0 until 0.26.0, the prompt field in CompletionRequest in /v1/completions accepts an unbounded list of strings or ints, and functions prompt_to_seq() and OnlineRenderer.preprocess_completion() expand every element, causing the server to create one engine generator and response slot per prompt. An authenticated API client can exhaust CPU, memory, async scheduling capacity, engine request slots, and response buffering with one request. This issue is fixed in version 0.26.0.

In vLLM prior to 0.27.0, an integer overflow in blockIdx.x * 2 * d in activation_kernels.cu can cause act_and_mul_kernel to consume another batched user's input, allowing a request processed in the same inference batch to receive a partial or complete copy of another user's inference result. This issue is fixed in version 0.27.0.

In vLLM from 0.20.2rc0 until 0.26.0, safe_load_prompt_embeds in vllm/renderers/embed_utils.py uses torch.sparse.check_sparse_tensor_invariants, whose process-global save, enable, and restore state can be raced by concurrent prompt_embeds parts submitted to POST /v1/chat/completions through AsyncMultiModalItemTracker.resolve_items, asyncio.gather, and the default executor, allowing an invalid sparse tensor to reach tensor.to_dense despite the CVE-2025-62164 guard when enable_prompt_embeds is enabled. This issue is fixed in version 0.26.0.

In vLLM prior to 0.26.0, the structured_outputs.regex parameter in vllm/v1/structured_output/backend_lm_format_enforcer.py is passed to lmformatenforcer.RegexParser without compile_regex_with_timeout or validation in validate_structured_output_request_lm_format_enforcer, allowing an unauthenticated /v1/completions request against the lm-format-enforcer backend to consume a CPU core and stall the structured-output engine path with a catastrophic regular expression. This issue is fixed in version 0.26.0.

In vLLM prior to 0.26.0, the validation_exception_handler in vllm/entrypoints/openai/server_utils.py converts FastAPI RequestValidationError objects with str(exc), and sanitize_message in vllm/entrypoints/utils.py does not remove traceback-style file paths, allowing unauthenticated malformed JSON requests to /v1/chat/completions, /v1/completions, /tokenize, and /detokenize to disclose the OS username, home and virtual-environment paths, Python version, internal package structure, line numbers, and endpoint handler names. This issue is fixed in version 0.26.0.