CISA list

● PyPI package

vllm

Vllm has 113 security advisories. The newest release any of them names as the fix is 0.32.0.

As of , vllm has 113 security advisories. The newest release any of them names as the fix is 0.32.0.

Your version, and the version to upgrade to

Find the row your installed version falls in. The second column is the first release after it that none of the advisories lists as affected; a release outside every row is one no advisory lists. That is what the advisories say, not a promise that a release is free of flaws.
If you runUpgrade toAdvisories
Before 0.32.00.32.0113 advisories

Security advisories, newest first

Security advisories, newest first
AdvisoryWhat it isAffectsFixed inPublished
CVE-2026-115913GHSA-8397-8qf6-f2gmvLLM: Phi-4 multimodal audio encoder materializes and retains full attention matrices: a single max-length audio OOM-crashes the vLLM engineModerate severityBefore 0.32.00.32.0
CVE-2026-115914GHSA-gq5q-89c7-vhvxvLLM DeepSeek-OCR image processing vulnerable to denial of serviceModerate severityFrom 0.11.1 before 0.32.00.32.0
CVE-2026-114053GHSA-h3rc-6mm3-gc2mvLLM: Request-controlled mm_processor_kwargs.code_revision allows remote code executionHigh severityFrom 0.7.3 before 0.31.00.31.0
CVE-2026-115918GHSA-x567-49g5-p4cf, PYSEC-2026-4349vLLM: Responses API bypasses trust_request_chat_template for chat_template_kwargsModerate severityFrom 0.16.0 before 0.32.00.32.0
CVE-2026-115917GHSA-vvpg-rrww-wpmh, PYSEC-2026-4348vLLM: Permissive default CORS origins allow cross-origin requests to the vLLM HTTP serverLow severityBefore 0.32.00.32.0
CVE-2026-115916GHSA-q97c-p368-75c4, PYSEC-2026-4347vLLM: Out-of-vocabulary prompt token id passes input validation and triggers a CUDA device-side assert (engine crash / DoS)Moderate severityBefore 0.32.00.32.0
CVE-2026-115915GHSA-h3rc-j82p-2c3h, PYSEC-2026-4346vLLM: Untrusted EPS inputs can trigger denial of service or Ghostscript RCEHigh severityBefore 0.32.00.32.0
CVE-2026-115912GHSA-6757-hggm-hp2w, PYSEC-2026-4345vLLM: Public multimodal SHM handle forgery reaches worker-side `pickle.loads()` with insecure serialization disabledHigh severityBefore 0.32.00.32.0
CVE-2026-115911GHSA-542m-9234-8gvf, PYSEC-2026-4344vLLM: DoS via unbounded chat_template Jinja2 rendering exhausts the renderer thread poolModerate severityFrom 0.19.0 before 0.32.00.32.0
CVE-2026-116178PYSEC-2026-4353vLLM from 0.20.0 before 0.30.0 contains a denial of service vulnerability that allows remote attackers to crash the engine by sending inconsistent multimodal metadata to the disaggregated /inference/v1/generate endpoint.From 0.20.0 before 0.30.00.30.0
CVE-2026-116177PYSEC-2026-4352vLLM from 0.10.2 before 0.25.0 contains an uncontrolled resource consumption vulnerability that allows unauthenticated attackers to exhaust memory because the Nemotron-VL processors globally set Image.MAXIMAGEPIXELS to None.From 0.10.2 before 0.25.00.25.0
CVE-2026-116176PYSEC-2026-4351vLLM from 0.9.2 before 0.24.0 contains a server-side request forgery vulnerability that allows API callers to force outbound connections by embedding a decode address in the X-Request-Id header.From 0.9.2 before 0.24.00.24.0
CVE-2026-116173PYSEC-2026-4350vLLM from 0.22.0 before 0.24.0 contains a denial of service vulnerability that allows remote attackers to terminate EngineCore by bypassing logprobtokenids limits in the Rust gRPC frontend.From 0.22.0 before 0.24.00.24.0
CVE-2026-114055GHSA-gx7p-2j49-hfq4vLLM : DoS via unvalidated mm_processor_kwargs in multimodal image processors (Isaac, InternVL, H2OVL, etc.)High severityFrom 0.7.3 before 0.31.00.31.0
CVE-2026-114050GHSA-p92p-rxj5-7p2xvLLM: Forged multimodal UUID + P0/P1 processor-cache eviction drift: unauthenticated remote crash of the whole vLLM engineModerate severityBefore 0.31.00.31.0
CVE-2026-114054PYSEC-2026-4362vLLM before 0.31.0 contains a cache aliasing vulnerability that allows authenticated attackers to collide shared cache keys by supplying caller-controlled values such as EXIF ImageID or cachesalt.Before 0.31.00.31.0
CVE-2026-114052PYSEC-2026-4361vLLM after 0.24.0 and before 0.31.0 contains a memory exhaustion vulnerability in MOSS-Audio model serving that caches a processor for every distinct request-controlled mmprocessorkwargs value in an unbounded dictionary.From 0.24.0 before 0.31.00.31.0
CVE-2026-105922PYSEC-2026-4319A security flaw has been discovered in vllm-project vLLM up to 0.31.0.Before 0.32.00.32.0
CVE-2026-105752GHSA-935w-9g4m-p28p, PYSEC-2026-4308vLLM: Harmony tool continuations drop `cache_salt`, restoring a cross-tenant prefix-cache membership oracleLow severityBefore 0.30.00.30.0
CVE-2026-105758GHSA-x6mc-67gf-chw4, PYSEC-2026-4310vLLM: Qwen2-VL / Qwen3-VL video samplers bound on request-controlled max_frames, which the num_frames ceiling does not reachModerate severityFrom 0.24.0 before 0.30.00.30.0
CVE-2026-105760GHSA-58v5-2m8f-94pr, PYSEC-2026-4307vLLM: GLMGA video sampling permits request-driven CPU and memory exhaustionModerate severityFrom 0.23.0rc2 before 0.30.00.30.0
CVE-2026-105754GHSA-ph72-cqr5-qpp7, PYSEC-2026-4309vLLM: Scale-out disaggregated multimodal transport trusts caller-supplied featuresModerate severityBefore 0.30.00.30.0
CVE-2026-105756GHSA-2823-qmq8-rwvj, PYSEC-2026-4306vLLM: Loose `cache_salt` validation lets a single request kill EngineCore on LMCache-MP deployments, uncaught downstream `ValueError` denial of serviceModerate severityBefore 0.30.00.30.0
CVE-2026-105759PYSEC-2026-4325vLLM is an inference and serving engine for large language models.Before 0.30.00.30.0
CVE-2026-105757GHSA-85xf-c7hm-whqw, PYSEC-2026-4199vLLM: Structured-output request errors escape the request boundary and terminate the shared EngineCore, engine-fatal denial of service (3 sites)Moderate severityBefore 0.30.00.30.0
CVE-2026-105755GHSA-2phq-3phc-84px, PYSEC-2026-4198vLLM: Flash late-interaction scoring caches query embeddings under a caller-controlled request id, cross-request integrity break and induced errors on `/score` and `/rerank`Moderate severityBefore 0.30.00.30.0
CVE-2026-105753GHSA-ph3r-5jfg-f84f, PYSEC-2026-4197vLLM: Mirrored multimodal IPC caches desync after a rejected request, a later request reusing the same media hash trips a receiver assertion in the engine coreModerate severityBefore 0.28.00.28.0
CVE-2026-103241PYSEC-2026-4186A flaw has been found in vllm-project vLLM up to 0.26.0.Before 0.27.00.27.0
CVE-2026-100653PYSEC-2026-4324vLLM is an inference and serving engine for large language models.From 0.22.1 before 0.28.00.28.0
CVE-2026-100652PYSEC-2026-4185vLLM versions 0.22.0 through 0.23.0 fail to validate stoptokenids against vocabulary bounds in Rust HTTP and gRPC frontends, allowing out-of-vocabulary token IDs to reach MinTokensLogitsProcessor.From 0.22.0 before 0.24.00.24.0
CVE-2026-100650PYSEC-2026-4184vLLM through 0.29.0 fetches and fully materializes remote or inline media before enforcing its documented media controls (the VLLMMAXAUDIOCLIPFILESIZEMB compressed-audio size cap, default 25 MB, and the per-modality --limit-mm-per-prompt...Before 0.30.00.30.0
CVE-2026-94626PYSEC-2026-4008vLLM through 0.29.0 fails to validate the tpsize parameter in kvtransferparams on OpenAI-compatible completion endpoints, allowing attackers to allocate unbounded memory.Before 0.30.00.30.0
CVE-2026-94625PYSEC-2026-4007vLLM through 0.29.0 contains a resource exhaustion vulnerability in MooncakeConnector where rejected prefill requests create ownerless transfer placeholders that are never reclaimed.Before 0.30.00.30.0
CVE-2026-94624GHSA-frvg-qw88-hww9, PYSEC-2026-4006vllm p2p never connected zmq_exhaustionModerate severityBefore 0.32.00.32.0
CVE-2026-94623PYSEC-2026-4005vLLM 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...Before 0.30.00.30.0
CVE-2026-94622PYSEC-2026-4004vLLM versions through 0.29.0 contain a denial of service vulnerability in the NIXL connector's metadata handling for prefill/decode disaggregated deployments.Before 0.30.00.30.0
CVE-2026-93989PYSEC-2026-4000vLLM through 0.29.0 fails to properly validate badwords token indices against the model's generation output width in SamplingParams.updatefromtokenizer().Before 0.30.00.30.0
CVE-2026-93841PYSEC-2026-3999vLLM through 0.29.0 contains a memory corruption vulnerability in the Triton bincountkernel where prompt token IDs index the penalty prompt-presence bitset without bounds checking against vocabulary size.Before 0.30.00.30.0
CVE-2026-93840PYSEC-2026-3998vLLM before 0.29.0 validates allowedtokenids against tokenizer length instead of model output logits width in SamplingParams.validateallowedtokenids().Before 0.29.00.29.0
CVE-2026-93592PYSEC-2026-3997vLLM 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.Before 0.28.00.28.0
CVE-2026-93436PYSEC-2026-3996vLLM through 0.29.0 fails to properly clean up decode-side metadata for rejected inference requests in prefill/decode disaggregated deployments.Before 0.30.00.30.0
CVE-2026-69147GHSA-8pw2-6jv3-mj5j, PYSEC-2026-4178vLLM: Request-selected PyNvVideoCodec GPU decode bypasses static VRAM reservationModerate severityBefore 0.28.00.28.0
CVE-2026-57173GHSA-hcwq-8wjf-3gcr, PYSEC-2026-4179vLLM: Unauthenticated audio decompression-bomb DoS in /v1/chat/completionsModerate severityBefore 0.24.00.24.0
CVE-2026-90553PYSEC-2026-3985vLLM before 0.28.0 contains a remote code execution vulnerability in the LlavaOnevision2 processor loader that ignores the trustremotecode parameter when loading remote processor classes.Before 0.28.00.28.0
CVE-2026-73560GHSA-4hhp-h66f-j5j7, PYSEC-2026-3934vLLM: SSRF + arbitrary local file read in MiMoV2OmniMultiModalProcessor `_fetch_image` and audio loader bypass MediaConnector protectionsModerate severityBefore 0.26.00.26.0
CVE-2026-73558GHSA-7m6h-x95x-82q5, PYSEC-2026-3935vLLM: Cross-User Data Leak VulnerabilityModerate severityBefore 0.27.00.27.0
CVE-2026-73557GHSA-pr7f-p5mw-fc87, PYSEC-2026-3938vLLM: Incomplete CVE-2025-62164 remediation can be bypassed by concurrent prompt partsModerate severityFrom 0.21.0 before 0.26.00.26.0
CVE-2026-73556GHSA-48jh-3gj7-fg8v, PYSEC-2026-3933vLLM: ReDoS via structured_outputs.regex in the lm-format-enforcer backend (no compile timeout), missed sibling of GHSA-rwxx-mrjm-wc2mModerate severityBefore 0.26.00.26.0
CVE-2026-73555GHSA-hwrm-c4cx-rf4j, PYSEC-2026-3937vLLM: Unauthenticated Internal Path and Username Disclosure via Validation Error MessagesModerate severityBefore 0.26.00.26.0
CVE-2026-71486GHSA-8737-qx52-hjff, PYSEC-2026-3936vLLM: Derender endpoints decode caller-supplied GenerateResponse token IDs without output boundsModerate severityBefore 0.26.00.26.0
CVE-2026-78684PYSEC-2026-4003vLLM before 0.27.0 fails to properly classify DeepStream as a GPU backend and omits pixel-limit enforcement in its decode path.Before 0.27.00.27.0
CVE-2026-73559GHSA-87x5-vmc3-756j, PYSEC-2026-3704vLLM: Completion prompt lists fan out into unbounded engine requestsModerate severityFrom 0.19.0 before 0.26.00.26.0
CVE-2026-54234GHSA-8wr5-jm2h-8r4f, PYSEC-2026-3542vLLM has Remote DoS via Invalid Recovered Token ReinjectionHigh severityFrom 0.17.1 before 0.24.00.24.0
CVE-2026-55574GHSA-rwxx-mrjm-wc2m, PYSEC-2026-2304vLLM: ReDoS via structured_outputs.regex compiled without timeout in xgrammar and outlines backendsHigh severityBefore 0.24.00.24.0
CVE-2026-55514GHSA-33cg-gxv8-3p8g, PYSEC-2026-2303vLLM denial of service via prompt embeds on M-RoPE modelsHigh severityFrom 0.12.0 before 0.24.00.24.0
CVE-2026-55646GHSA-v82g-2437-67m2, PYSEC-2026-2305vLLM: Speech-to-text upload size limit is enforced after full UploadFile readModerate severityFrom 0.22.0 before 0.24.00.24.0
CVE-2026-54232PYSEC-2026-227vLLM is an inference and serving engine for large language models (LLMs).Before 0.22.10.22.1
CVE-2026-54233GHSA-6pr9-rp53-2pmc, PYSEC-2026-3404vLLM: OOM Denial of Service via Audio Decompression BombModerate severityBefore 0.24.00.24.0
CVE-2026-54236GHSA-hgg8-fqqc-vfmw, PYSEC-2026-3408vLLM: incomplete CVE-2026-22778 fix leaks PIL repr addresses via Anthropic routerModerate severityBefore 0.24.00.24.0
CVE-2026-53923GHSA-5jv2-g5wq-cmr4, PYSEC-2026-3403vLLM: GGUF dequantize kernel int truncation exposes uninitialized GPU memory in multi-tenant servingModerate severityFrom 0.5.5 before 0.24.00.24.0
CVE-2026-12491GHSA-8jr5-v98p-w75m, PYSEC-2026-3406vLLM: image EXIF Rotation & PNG tRNS Transparency Not Normalized, Causing Mismatch Between Model Input and ExpectationsModerate severityFrom 0.11.0 before 0.24.00.24.0
CVE-2026-54235GHSA-7h4p-rffg-7823, PYSEC-2026-3405vLLM: temperature=NaN and temperature=Infinity bypass validation and propagate to GPU kernelsModerate severityBefore 0.24.00.24.0
CVE-2026-48746GHSA-94f4-hr76-p5j6, PYSEC-2026-226vLLM: OpenAI auth bypassCritical severityFrom 0.3.0 before 0.22.00.22.0
CVE-2026-41523GHSA-q8gq-377p-jq3r, PYSEC-2026-2300vLLM: Security Check Bypass via assert Statement in Activation Function Loading Allows Arbitrary Code ExecutionHigh severityBefore 0.22.00.22.0
CVE-2026-5497GHSA-wcwg-c5fc-9vrc, PYSEC-2026-2302vLLM is vulnerable to an Out-of-Memory (OOM) Denial of Service (DoS) attack due to unbounded frame count processing in the `VideoMediaIO.load_base64()` methodHigh severityFrom 0.8.0 before 0.19.00.19.0
CVE-2026-47155GHSA-3ww4-5jv9-j5gm, PYSEC-2026-2301vLLM's Artifact Pin Decay allows pinned deployments to load unpinned code, weights, and processorsModerate severityBefore 0.22.00.22.0
CVE-2026-9540GHSA-98f3-hwg4-4rf7, PYSEC-2026-3407vllm has Improper Resource Shutdown or ReleaseModerate severityUp to and including 0.19.0No fixed version listed
CVE-2026-44223GHSA-83vm-p52w-f9pw, PYSEC-2026-145vLLM: extract_hidden_states speculative decoding crashes server on any request with penalty parametersModerate severityFrom 0.18.0 before 0.20.00.20.0
CVE-2026-44222GHSA-hpv8-x276-m59f, PYSEC-2026-3409vLLM Vulnerable to Remote DoS via Special-Token PlaceholdersModerate severityFrom 0.6.1 before 0.20.00.20.0
CVE-2026-7141GHSA-x368-4g9h-fvv4, PYSEC-2026-2306vLLM makes Use of Uninitialized ResourceLow severityBefore 0.19.10.19.1
CVE-2026-34755GHSA-pq5c-rjhq-qp7p, PYSEC-2026-144vLLM: Denial of Service via Unbounded Frame Count in video/jpeg Base64 ProcessingModerate severityFrom 0.7.0 before 0.19.00.19.0
CVE-2026-34753GHSA-pf3h-qjgv-vcpr, PYSEC-2026-3410vLLM: Server-Side Request Forgery (SSRF) in `download_bytes_from_url `Moderate severityFrom 0.16.0 before 0.19.00.19.0
CVE-2026-34756GHSA-3mwp-wvh9-7528, PYSEC-2026-2298vLLM: Unauthenticated OOM Denial of Service via Unbounded `n` Parameter in OpenAI API ServerModerate severityFrom 0.1.0 before 0.19.00.19.0
CVE-2026-34760GHSA-6c4r-fmh3-7rh8, PYSEC-2026-2299vLLM: Processing differential in multi-channel audio downmixing enables hidden-input/moderation bypass for audio modelsModerate severityFrom 0.5.5 before 0.18.00.18.0
CVE-2026-27893GHSA-7972-pg2x-xr59, PYSEC-2026-2297vLLM has Hardcoded Trust Override in Model Files Enables RCE Despite Explicit User Opt-OutHigh severityFrom 0.10.1 before 0.18.00.18.0
CVE-2026-25960GHSA-v359-jj2v-j536, PYSEC-2026-3411vLLM has SSRF Protection BypassModerate severityFrom 0.15.1 before 0.17.00.17.0
CVE-2026-22778GHSA-4r2x-xpjr-7cvv, PYSEC-2026-565vLLM has RCE In Video ProcessingCritical severityFrom 0.8.3 before 0.14.10.14.1
CVE-2026-24779GHSA-qh4c-xf7m-gxfc, PYSEC-2026-2020vLLM vulnerable to Server-Side Request Forgery (SSRF) through MediaConnectorHigh severityBefore 0.14.10.14.1
CVE-2026-22807GHSA-2pc9-4j83-qjmr, PYSEC-2026-2010vLLM affected by RCE via auto_map dynamic module loading during model initializationHigh severityFrom 0.10.1 before 0.14.00.14.0
CVE-2026-22773GHSA-grg2-63fw-f2qr, PYSEC-2026-143vLLM is vulnerable to DoS in Idefics3 vision models via image payload with ambiguous dimensionsModerate severityFrom 0.6.4 before 0.12.00.12.0
CVE-2026-56340GHSA-mcmc-2m55-j8jj, PYSEC-2026-250vLLM introduced enhanced protection for CVE-2025-62164High severityFrom 0.10.2 before 0.13.00.13.0
CVE-2025-66448GHSA-8fr4-5q9j-m8gm, PYSEC-2026-2015vLLM vulnerable to remote code execution via transformers_utils/get_configHigh severityBefore 0.11.10.11.1
CVE-2025-62426GHSA-69j4-grxj-j64p, PYSEC-2026-2012vLLM vulnerable to DoS via large Chat Completion or Tokenization requests with specially crafted `chat_template_kwargs`Moderate severityFrom 0.5.5 before 0.11.10.11.1
CVE-2025-62372GHSA-pmqf-x6x8-p7qw, PYSEC-2026-2019vLLM vulnerable to DoS with incorrect shape of multimodal embedding inputsHigh severityFrom 0.5.5 before 0.11.10.11.1
CVE-2025-62164GHSA-mrw7-hf4f-83pf, PYSEC-2026-2018vLLM deserialization vulnerability leading to DoS and potential RCEHigh severityFrom 0.10.2 before 0.11.10.11.1
CVE-2025-6242GHSA-3f6c-7fw2-ppm4, PYSEC-2026-2011vLLM is vulnerable to Server-Side Request Forgery (SSRF) through `MediaConnector` classHigh severityFrom 0.5.0 before 0.11.00.11.0
CVE-2025-61620GHSA-6fvq-23cw-5628, PYSEC-2026-2013vLLM: Resource-Exhaustion (DoS) through Malicious Jinja Template in OpenAI-Compatible ServerModerate severityFrom 0.5.1 before 0.11.00.11.0
CVE-2025-59425GHSA-wr9h-g72x-mwhm, PYSEC-2026-2026vLLM is vulnerable to timing attack at bearer authHigh severityBefore 0.11.00.11.0
CVE-2025-9141GHSA-79j6-g2m3-jgfw, PYSEC-2026-2014vLLM has remote code execution vulnerability in the tool call parser for Qwen3-CoderHigh severityFrom 0.10.0 before 0.10.1.10.10.1.1
CVE-2025-48956GHSA-rxc4-3w6r-4v47, PYSEC-2026-2021vllm API endpoints vulnerable to Denial of Service AttacksHigh severityFrom 0.1.0 before 0.10.1.10.10.1.1
CVE-2025-48944GHSA-vrq3-r879-7m65, PYSEC-2026-2023vLLM Tool Schema allows DoS via Malformed pattern and type FieldsModerate severityFrom 0.8.0 before 0.9.00.9.0
CVE-2025-48943GHSA-9hcf-v7m4-6m2j, PYSEC-2025-55vLLM allows clients to crash the openai server with invalid regexModerate severityFrom 0.8.0 before 0.9.00.9.0
CVE-2025-48942GHSA-6qc9-v4r8-22xg, PYSEC-2025-54vLLM DOS: Remotely kill vllm over http with invalid JSON schemaModerate severityFrom 0.8.0 before 0.9.00.9.0
CVE-2025-46722GHSA-c65p-x677-fgj6, PYSEC-2025-43vLLM has a Weakness in MultiModalHasher Image Hashing ImplementationModerate severityBefore 0.9.00.9.0
CVE-2025-46570GHSA-4qjh-9fv9-r85r, PYSEC-2025-53Potential Timing Side-Channel Vulnerability in vLLM’s Chunk-Based Prefix CachingLow severityBefore 0.9.00.9.0
CVE-2025-71379GHSA-j828-28rj-hfhpvLLM vulnerable to Regular Expression Denial of ServiceModerate severityFrom 0.6.3 before 0.9.00.9.0
CVE-2025-48887GHSA-w6q7-j642-7c25, PYSEC-2025-50vLLM has a Regular Expression Denial of Service (ReDoS, Exponential Complexity) Vulnerability in `pythonic_tool_parser.py`Moderate severityFrom 0.6.4 before 0.9.00.9.0
CVE-2025-47277GHSA-hjq4-87xh-g4fv, PYSEC-2026-567vLLM Allows Remote Code Execution via PyNcclPipe Communication ServiceCritical severityFrom 0.6.5 before 0.8.50.8.5
CVE-2025-30165GHSA-9pcc-gvx5-r5wm, PYSEC-2026-2017Remote Code Execution Vulnerability in vLLM Multi-Node Cluster ConfigurationHigh severityFrom 0.5.2 before 0.10.00.10.0
CVE-2025-46560GHSA-vc6m-hm49-g9qg, PYSEC-2026-2022vLLM: Quadratic Time Complexity in Input Token Processing leads to denial of serviceModerate severityFrom 0.8.0 before 0.8.50.8.5

The next 13 advisories

Other PyPI packages

Sources
  • Package advisories from the GitHub Advisory Database (CC BY 4.0). Version ranges are reworded as sentences (source). Licence: CC BY 4.0
  • Package advisories from the PyPA advisory database (CC BY 4.0). Version ranges are reworded as sentences (source). Licence: CC BY 4.0