Products / vllm
vllm
89 records.
- vLLM: DoS via unbounded chat_template Jinja2 rendering exhausts the renderer thread pool A single short chat template with unbounded Jinja2 rendering runs in a limited worker pool that cannot be interrupted, blocking chat endpoints for all tenants.
- vLLM: Public multimodal SHM handle forgery reaches worker-side `pickle.loads()` with insecure serialization disabled Forged shared-memory handles in vLLM's multimodal cache reach pickle.loads in the worker, and this works without the insecure-serialisation opt-in.
- vLLM: Phi-4 multimodal audio encoder materializes and retains full attention matrices: a single max-length audio OOM-crashes the vLLM engine A legal maximum-length audio clip makes the Phi-4 audio encoder keep full attention matrices for every layer in memory, exhausting GPU memory and crashing the engine.
- vLLM DeepSeek-OCR image processing vulnerable to denial of service A valid image whose normalised pixels sum to exactly zero hits an inherited early-return guard in the DeepSeek-OCR path, crashing the serving process.
- vLLM: Untrusted EPS inputs can trigger denial of service or Ghostscript RCE Image bytes from a request are autodetected by Pillow without a format allowlist, so an EPS file labeled as PNG reaches Ghostscript and runs commands.
- vLLM: Out-of-vocabulary prompt token id passes input validation and triggers a CUDA device-side assert (engine crash / DoS) A pre-tokenized prompt with an id one past the embedding range passes input validation and crashes the inference engine's CUDA context.
- vLLM: Permissive default CORS origins allow cross-origin requests to the vLLM HTTP server The server defaults to a wildcard CORS origin and has no auth unless an API key is set, so any web page can read responses from a reachable instance.
- vLLM: Responses API bypasses trust_request_chat_template for chat_template_kwargs The Responses endpoint accepts request-supplied Jinja chat templates that skip the trust-flag check, exhausting CPU and memory.
- vllm p2p never connected zmq_exhaustion Repeated API requests naming many unreachable peer addresses leave ZeroMQ sockets open, exhausting the socket quota and crashing the inference engine.
- vLLM: Request-controlled mm_processor_kwargs.code_revision allows remote code execution A client request field selects another repository revision for processor code, so Transformers imports attacker-chosen Python when trust_remote_code is enabled.
- vLLM : DoS via unvalidated mm_processor_kwargs in multimodal image processors (Isaac, InternVL, H2OVL, etc.) Clients can set per-request processor arguments such as image size that are passed unchecked to processors, forcing huge allocations that exhaust server memory.
- vLLM: Forged multimodal UUID + P0/P1 processor-cache eviction drift: unauthenticated remote crash of the whole vLLM engine Attacker-chosen multimodal UUIDs desynchronize two processor caches, so a crafted image request crashes the whole inference engine without authentication.
- vLLM: Harmony tool continuations drop `cache_salt` — restoring a cross-tenant prefix-cache membership oracle Tool continuation turns are cached without the tenant's cache salt, so another tenant can infer guessed prompts from exact cached-token counts.
- vLLM: Mirrored multimodal IPC caches desync after a rejected request — a later request reusing the same media hash trips a receiver assertion in the engine core A rejected request leaves the frontend cache believing media is stored in the engine, so a later request reusing that media hash triggers a fatal assertion.
- vLLM: Loose `cache_salt` validation lets a single request kill EngineCore on LMCache-MP deployments — uncaught downstream `ValueError` denial of service One request with a cache_salt containing a forbidden character passes loose validation and raises an uncaught error that kills the vLLM engine process.
- vLLM: Flash late-interaction scoring caches query embeddings under a caller-controlled request id — cross-request integrity break and induced errors on `/score` and `/rerank` A caller-chosen request ID lets a second request overwrite the cached query embedding, so documents are scored against the attacker's query.
- vLLM: GLMGA video sampling permits request-driven CPU and memory exhaustion Chat requests can set video sampling options that make the server build a huge frame index, consuming CPU and memory and delaying other media requests.
- vLLM: Structured-output request errors escape the request boundary and terminate the shared EngineCore — engine-fatal denial of service (3 sites) Malformed structured-output requests raise an uncaught error in the engine loop, which kills the shared engine and blocks all other requests.
- vLLM: Scale-out disaggregated multimodal transport trusts caller-supplied features The generate endpoint accepts internal multimodal fields from callers as if they came from the trusted renderer, so forged values can crash the shared engine process.
- vLLM: Qwen2-VL / Qwen3-VL video samplers bound on request-controlled max_frames, which the num_frames ceiling does not reach Unauthenticated requests set uncapped frame and fps limits for Qwen vision models, inflating API server memory use.
- vLLM: Request-selected PyNvVideoCodec GPU decode bypasses static VRAM reservation A client can pick a GPU video decoder per request, allocating GPU memory the engine never reserved and exhausting VRAM.
- vLLM: Unauthenticated audio decompression-bomb DoS in /v1/chat/completions An unauthenticated client sends a small compressed audio clip to the chat endpoint, which expands to gigabytes on decode because the duration limit applies only to transcription.
- vLLM: SSRF + arbitrary local file read in MiMoV2OmniMultiModalProcessor `_fetch_image` and audio loader bypass MediaConnector protections A multimodal processor fetches user-supplied image and audio URLs and local paths directly, skipping the media connector's SSRF and local path checks.
- vLLM: Cross-User Data Leak Vulnerability An integer overflow in an activation kernel's index arithmetic makes some batched requests read and return another user's inference output.
- vLLM: ReDoS via structured_outputs.regex in the lm-format-enforcer backend (no compile timeout) — missed sibling of GHSA-rwxx-mrjm-wc2m A structured output request with a catastrophic regex is compiled without a timeout in the lm-format-enforcer backend, stalling the inference worker.
- vLLM: Derender endpoints decode caller-supplied GenerateResponse token IDs without output bounds Derender routes detokenize client-supplied token ID lists without output size or count limits, consuming CPU and memory.
- vLLM: Unauthenticated Internal Path and Username Disclosure via Validation Error Messages Malformed API requests trigger validation errors whose text includes server file paths, line numbers and handler names, which unauthenticated callers can read.
- vLLM: Incomplete CVE-2025-62164 remediation can be bypassed by concurrent prompt parts Two concurrent prompt-embedding parts race on a process-wide sparse-check flag, letting a malformed sparse tensor reach dense conversion unchecked.
- vLLM: Completion prompt lists fan out into unbounded engine requests A completions request with a list of prompts is expanded into one engine generator per element with no outer count limit, so one request spawns many backend subrequests.
- vLLM denial of service via prompt embeds on M-RoPE models A prompt-embeds request to completions on an M-RoPE model trips an assertion that crashes the engine and the whole server.
- vLLM: Processing differential in multi-channel audio downmixing enables hidden-input/moderation bypass for audio models Content hidden in unused audio channels is heard by the model after mono downmixing but not by human listeners, evading moderation.
- vLLM has Remote DoS via Invalid Recovered Token Reinjection An overlapping generate and abort sequence makes speculative decoding feed an out-of-vocabulary sentinel token back into the drafter, crashing the shared worker.
- vLLM: ReDoS via structured_outputs.regex compiled without timeout in xgrammar and outlines backends A user-supplied regex for structured output is compiled by grammar backends with no timeout, so a catastrophic pattern hangs the inference worker.
- vLLM: Speech-to-text upload size limit is enforced after full UploadFile read The speech-to-text upload handlers read the whole file before the documented size limit is checked, so oversized uploads consume memory first.
- vLLM: GGUF dequantize kernel int truncation exposes uninitialized GPU memory in multi-tenant serving Integer truncation in GGUF dequantization leaves part of a newly allocated GPU buffer unwritten, exposing residual memory from earlier requests.
- vLLM: OOM Denial of Service via Audio Decompression Bomb A small compressed audio upload decodes into a very large float buffer, exhausting server memory because the size check only applies to compressed bytes.
- vLLM: temperature=NaN and temperature=Infinity bypass validation and propagate to GPU kernels Temperature checks use comparisons that let NaN and positive infinity through, so these values reach GPU sampling kernels and crash the inference worker.
- vLLM: image EXIF Rotation & PNG tRNS Transparency Not Normalized, Causing Mismatch Between Model Input and Expectations Images with EXIF rotation or PNG transparency tricks are preprocessed differently than humans see them, so the model receives hidden or altered content.
- vLLM: incomplete CVE-2026-22778 fix leaks PIL repr addresses via Anthropic router Malformed image bytes make the Anthropic-compatible router echo an unsanitised exception containing a heap address, because a later endpoint skipped the earlier fix.
- vLLM: OpenAI auth bypass A crafted Host header changes the path the auth middleware parses, so /v1 requests skip the API key check while still being routed.
- vLLM: Security Check Bypass via assert Statement in Activation Function Loading Allows Arbitrary Code Execution The only check on which activation function a model config may load is an assert, which Python strips in optimised mode, so a malicious config runs arbitrary code.
- vLLM is vulnerable to an Out-of-Memory (OOM) Denial of Service (DoS) attack due to unbounded frame count processing in the `VideoMediaIO.load_base64()` method A single unauthenticated chat completion request with a data URL containing thousands of comma-separated JPEG frames forces the server to decode them all into memory.
- vLLM's Artifact Pin Decay allows pinned deployments to load unpinned code, weights, and processors Revision pins apply to only some artifacts, so dynamic code, processors and side weights can be loaded from an unpinned revision.
- vllm has Improper Resource Shutdown or Release Requests to vLLM's OpenAI-compatible serving path trigger a resource release flaw that remote attackers can use to take the server down.
- vLLM: extract_hidden_states speculative decoding crashes server on any request with penalty parameters A speculative decoding proposer returns a wrongly shaped tensor after the first step, so any request with penalty settings crashes the engine process.
- vLLM Vulnerable to Remote DoS via Special-Token Placeholders Unauthenticated prompts spelling special placeholder tokens without matching media cause an out-of-range index that crashes the worker.
- vLLM makes Use of Uninitialized Resource A helper in vLLM's KV cache interface reads an uninitialised resource, reachable remotely though hard to exploit.
- vLLM: Unauthenticated OOM Denial of Service via Unbounded `n` Parameter in OpenAI API Server An unauthenticated request with an enormous n value makes the API server copy request objects until it blocks the event loop and runs out of memory.
- vLLM: Server-Side Request Forgery (SSRF) in `download_bytes_from_url ` The batch runner fetches any URL supplied in batch input JSON without validation, letting callers reach internal hosts and metadata endpoints.
- vLLM: Denial of Service via Unbounded Frame Count in video/jpeg Base64 Processing The base64 video/jpeg path splits frames without applying the frame limit, so one request with thousands of frames exhausts server memory.