vLLM <= 0.23.0 - Anthropic Router Heap Address Information Leak

vLLM = 0.23.0 incompletely fixes CVE-2026-22778. The original fix added sanitizemessage to the OpenAI router but the Anthropic-compatible router /v1/messages echoes strexc directly. id: CVE-2026-54236 info: name: vLLM = 0.23.0 - Anthropic Router Heap Address Information Leak author: kenlacroix...

CVE-2026-73560

vLLM is an inference and serving engine for large language models. Prior to 0.26.0, the MiMoV2OmniMultiModalProcessor in vllm/transformersutils/processors/mimov2omni.py passes attacker-controlled image and audio strings through fetchimage, requests.get, and Image.open instead of MediaConnector,...

CVE-2026-73560

vLLM is an inference and serving engine for large language models. Prior to 0.26.0, the MiMoV2OmniMultiModalProcessor in vllm/transformersutils/processors/mimov2omni.py passes attacker-controlled image and audio strings through fetchimage, requests.get, and Image.open instead of MediaConnector,...

EUVD-2026-60377

vLLM is an inference and serving engine for large language models. Prior to 0.26.0, the MiMoV2OmniMultiModalProcessor in vllm/transformersutils/processors/mimov2omni.py passes attacker-controlled image and audio strings through fetchimage, requests.get, and Image.open instead of MediaConnector,...

CVE-2026-73560

vLLM (an LLM inference and serving engine) prior to 0.26.0 is vulnerable to SSRF and arbitrary local file read via the MiMoV2OmniMultiModalProcessor in vllm/transformers_utils/processors/mimo_v2_omni.py. The _fetch_image function and audio loader pass attacker-controlled image/audio strings direc...

CVE-2026-73560 vLLM: SSRF + arbitrary local file read in MiMoV2OmniMultiModalProcessor `_fetch_image` and audio loader bypass MediaConnector protections

vLLM is an inference and serving engine for large language models. Prior to 0.26.0, the MiMoV2OmniMultiModalProcessor in vllm/transformersutils/processors/mimov2omni.py passes attacker-controlled image and audio strings through fetchimage, requests.get, and Image.open instead of MediaConnector,...

CVE-2026-73560 vLLM: SSRF + arbitrary local file read in MiMoV2OmniMultiModalProcessor `_fetch_image` and audio loader bypass MediaConnector protections

vLLM is an inference and serving engine for large language models. Prior to 0.26.0, the MiMoV2OmniMultiModalProcessor in vllm/transformersutils/processors/mimov2omni.py passes attacker-controlled image and audio strings through fetchimage, requests.get, and Image.open instead of MediaConnector,...

CVE-2026-71486

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 generateresponses, choices, tokenids, promptlogprobs, logprobs.content, toplogprobs, a...

CVE-2026-71486 vLLM: Derender endpoints decode caller-supplied GenerateResponse token IDs without output bounds

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 generateresponses, choices, tokenids, promptlogprobs, logprobs.content, toplogprobs, a...

CVE-2026-71486

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 generateresponses, choices, tokenids, promptlogprobs, logprobs.content, toplogprobs, a...

CVE-2026-71486

In vLLM versions prior to 0.26.0, the /v1/completions/derender and /v1/chat/completions/derender endpoints process caller-supplied GenerateResponse objects via OnlineDerenderer and tokenizer.decode before enforcing max_model_len, max_tokens, max_num_seqs, or response-size limits. An authenticated...

CVE-2026-71486 vLLM: Derender endpoints decode caller-supplied GenerateResponse token IDs without output bounds

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 generateresponses, choices, tokenids, promptlogprobs, logprobs.content, toplogprobs, a...

CVE-2026-73555

A flaw was found in vLLM, an inference and serving engine for large language models. Unauthenticated attackers can send malformed JSON requests to various API endpoints, triggering validation errors. The system's error handling mechanism converts these validation errors into strings without...

CVE-2026-73557

A flaw was found in vLLM, an inference and serving engine for large language models. An incomplete remediation for a previous vulnerability CVE-2025-62164 can be bypassed. This occurs when concurrent prompt parts are processed, leading to a race condition in the handling of sparse tensors. This...

CVE-2026-73558

A flaw was found in vLLM, an inference and serving engine for large language models. An integer overflow vulnerability in the actandmulkernel kernel allows a remote attacker to cause cross-user data leakage. By crafting specific requests processed in the same inference batch, an attacker can obta...

CVE-2026-73559

A flaw was found in vLLM, an inference and serving engine for large language models. An authenticated API client can exploit this vulnerability by sending a single request to the /v1/completions endpoint with an excessively large list of prompts. This unbounded input causes the vLLM server to...

CVE-2026-73556

A flaw was found in vLLM, an inference and serving engine for large language models. An unauthenticated remote attacker can exploit a Regular Expression Denial of Service ReDoS vulnerability in the lm-format-enforcer backend. By submitting a specially crafted regular expression through the...

vLLM: Completion prompt lists fan out into unbounded engine requests

Summary The /v1/completions request model accepts prompt as a list of text prompts or a list of token-id prompts without any outer prompt-count bound. The serving path turns each element into a separate engine input, creates one engine generator per element, merges all generators, and allocates a...

GHSA-87X5-VMC3-756J vLLM: Completion prompt lists fan out into unbounded engine requests

Summary The /v1/completions request model accepts prompt as a list of text prompts or a list of token-id prompts without any outer prompt-count bound. The serving path turns each element into a separate engine input, creates one engine generator per element, merges all generators, and allocates a...

CVE-2026-73559

vLLM is an inference and serving engine for large language models. From 0.19.0 until 0.26.0, the /v1/completions CompletionRequest.prompt field in vllm/entrypoints/openai/completion/protocol.py accepts an unbounded liststr or listlistint, prompttoseq in vllm/renderers/inputs/preprocess.py and...