SB2026070221 - Multiple vulnerabilities in vLLM



SB2026070221 - Multiple vulnerabilities in vLLM

Published: July 2, 2026 Updated: August 25, 2026

Security Bulletin ID SB2026070221
CSH Severity
Medium
Patch available
YES
Number of vulnerabilities 12
Exploitation vector Remote access
Highest impact Denial of service

Breakdown by Severity

Medium 75% Low 25%
  • Low
  • Medium
  • High
  • Critical

Description

This security bulletin contains information about 12 vulnerabilities.


1) Reachable assertion (CVE-ID: CVE-2026-55514)

CWE-ID: CWE-617 - Reachable Assertion

CVSSv4: 7.1 [CVSS:4.0/AV:N/AC:L/AT:N/PR:L/UI:N/VC:N/VI:N/VA:H/SC:N/SI:N/SA:N]


The vulnerability allows a remote attacker to perform a denial of service (DoS) attack.

The vulnerability exists due to a reachable assertion. A remote user can send a pure prompt embeds payload in a "/v1/completions" request with a model using M-RoPE and cause a denial of service condition on the target system.


2) Allocation of Resources Without Limits or Throttling (CVE-ID: CVE-2026-55646)

CWE-ID: CWE-770 - Allocation of Resources Without Limits or Throttling

CVSSv4: 7.1 [CVSS:4.0/AV:N/AC:L/AT:N/PR:L/UI:N/VC:N/VI:N/VA:H/SC:N/SI:N/SA:N]


The vulnerability allows a remote user to cause a denial of service.

The vulnerability exists due to uncontrolled resource consumption in the speech-to-text upload handling routes when processing oversized audio file uploads. A remote user can send a specially crafted oversized multipart upload to cause a denial of service.

Exploitation requires that the deployment exposes the speech-to-text endpoints and has a speech-to-text capable model or task configured.


3) Input validation error (CVE-ID: CVE-2026-54234)

CWE-ID: CWE-20 - Improper input validation

CVSSv4: 8.7 [CVSS:4.0/AV:N/AC:L/AT:N/PR:N/UI:N/VC:N/VI:N/VA:H/SC:N/SI:N/SA:N]


The vulnerability allows a remote attacker to cause a denial of service.

The vulnerability exists due to improper input validation in the speculative decoding path when handling overlapping public gRPC Generate and Abort requests. A remote attacker can send a specific overlapping request sequence to cause a denial of service.

In shared deployments, exploitation can crash the engine worker, abort concurrent requests, and prevent later requests from completing until the worker is restarted. The issue was reproduced on Qwen3 GPTQ.


4) Inefficient regular expression complexity (CVE-ID: CVE-2026-55574)

CWE-ID: CWE-1333 - Inefficient Regular Expression Complexity

CVSSv4: 8.7 [CVSS:4.0/AV:N/AC:L/AT:N/PR:N/UI:N/VC:N/VI:N/VA:H/SC:N/SI:N/SA:N]


The vulnerability allows a remote attacker to cause a denial of service.

The vulnerability exists due to inefficient regular expression complexity in the structured_outputs.regex API in the xgrammar and outlines backends when compiling a user-supplied regex string. A remote attacker can send a specially crafted regex pattern to cause a denial of service.

A single crafted request can hang an inference worker indefinitely.


5) Improper handling of highly compressed data (CVE-ID: CVE-2026-54233)

CWE-ID: CWE-409 - Improper Handling of Highly Compressed Data (Data Amplification)

CVSSv4: 7.1 [CVSS:4.0/AV:N/AC:L/AT:N/PR:L/UI:N/VC:N/VI:N/VA:H/SC:N/SI:N/SA:N]


The vulnerability allows a remote user to cause a denial of service.

The vulnerability exists due to improper handling of highly compressed data in the /v1/audio/transcriptions endpoint when processing compressed audio uploads. A remote user can send a specially crafted audio file to cause a denial of service.

The issue arises because the endpoint limits compressed upload size but not decoded PCM output, allowing excessive memory consumption during audio decoding and concatenation.


6) Inclusion of Sensitive Information in Log Files (CVE-ID: CVE-2026-54236)

CWE-ID: CWE-532 - Information Exposure Through Log Files

CVSSv4: 6.9 [CVSS:4.0/AV:N/AC:L/AT:N/PR:N/UI:N/VC:L/VI:N/VA:N/SC:N/SI:N/SA:N]


The vulnerability allows a remote attacker to disclose sensitive information.

The vulnerability exists due to insertion of sensitive information into an error message in the Anthropic API router and realtime speech-to-text error-handling paths when processing malformed image data or other input that triggers exceptions containing object representations. A remote attacker can send crafted requests or input that causes exceptions to be returned with unsanitized error messages to disclose sensitive information.

The issue leaks object representation memory addresses in response bodies, including through in-route HTTP exception handling, an SSE streaming converter, and WebSocket error paths that bypass the sanitizing exception handler chain.


7) Improper Validation of Specified Type of Input (CVE-ID: CVE-2026-54235)

CWE-ID: CWE-1287 - Improper Validation of Specified Type of Input

CVSSv4: 8.7 [CVSS:4.0/AV:N/AC:L/AT:N/PR:N/UI:N/VC:N/VI:N/VA:H/SC:N/SI:N/SA:N]


The vulnerability allows a remote attacker to cause a denial of service.

The vulnerability exists due to improper validation of specified type of input in sampling_params.py when processing temperature values. A remote attacker can supply a temperature value of NaN or positive infinity to cause a denial of service.

The issue occurs because Python IEEE 754 float comparison semantics allow these non-finite values to bypass the validation checks and propagate to GPU sampling kernels, which can trigger undefined behavior or CUDA errors.


8) Incorrect Conversion between Numeric Types (CVE-ID: CVE-2026-53923)

CWE-ID: CWE-681 - Incorrect Conversion between Numeric Types

CVSSv4: 4.6 [CVSS:4.0/AV:L/AC:L/AT:N/PR:N/UI:A/VC:L/VI:L/VA:N/SC:N/SI:N/SA:N]


The vulnerability allows a remote attacker to disclose sensitive information.

The vulnerability exists due to incorrect conversion between numeric types in GGUF dequantize kernels in csrc/quantization/gguf/gguf_kernel.cu when processing a crafted GGUF model with tensor dimensions whose product exceeds INT_MAX. A remote attacker can supply a specially crafted GGUF model file to disclose sensitive information.

User interaction is required to load the crafted model for inference serving, and the issue can expose residual GPU memory from prior requests in multi-tenant deployments.


9) Input validation error (CVE-ID: N/A)

CWE-ID: CWE-20 - Improper input validation

CVSSv4: 6.3 [CVSS:4.0/AV:N/AC:L/AT:P/PR:N/UI:N/VC:N/VI:L/VA:L/SC:N/SI:N/SA:N]


The vulnerability allows a remote attacker to alter downstream model reasoning.

The vulnerability exists due to improper input validation in the image processing logic in vllm/multimodal/image.py when parsing images with EXIF orientation metadata. A remote attacker can supply a specially crafted image to alter downstream model reasoning.

The issue arises because EXIF orientation is not normalized before the image is processed by the model.


10) Input validation error (CVE-ID: N/A)

CWE-ID: CWE-20 - Improper input validation

CVSSv4: 6.3 [CVSS:4.0/AV:N/AC:L/AT:P/PR:N/UI:N/VC:N/VI:L/VA:L/SC:N/SI:N/SA:N]


The vulnerability allows a remote attacker to alter downstream model reasoning.

The vulnerability exists due to improper input validation in the image mode conversion logic in vllm/multimodal/image.py when parsing PNG images with tRNS transparency in non-RGBA modes. A remote attacker can supply a specially crafted image to alter downstream model reasoning.

Only the RGBA-to-RGB path is explicitly flattened; PNG images in P, L, or RGB modes with tRNS transparency fall back to direct RGB conversion, which can make subtle overlay elements visible.


11) Input validation error (CVE-ID: N/A)

CWE-ID: CWE-20 - Improper input validation

CVSSv4: 6.3 [CVSS:4.0/AV:N/AC:L/AT:P/PR:N/UI:N/VC:N/VI:L/VA:L/SC:N/SI:N/SA:N]


The vulnerability allows a remote attacker to alter downstream model reasoning.

The vulnerability exists due to improper input validation in the image loading logic in vllm/multimodal/image.py when parsing APNG or GIF files. A remote attacker can supply a specially crafted image to alter downstream model reasoning.

Only the first frame is loaded when processing APNG or GIF images.


12) Allocation of Resources Without Limits or Throttling (CVE-ID: CVE-2026-57173)

CWE-ID: CWE-770 - Allocation of Resources Without Limits or Throttling

CVSSv4: 7.1 [CVSS:4.0/AV:N/AC:L/AT:N/PR:L/UI:N/VC:N/VI:N/VA:H/SC:N/SI:N/SA:N]


The vulnerability allows a remote user to cause a denial of service.

The vulnerability exists due to allocation of resources without limits or throttling in the /v1/chat/completions chat audio processing path when decoding compressed audio input. A remote user can submit a specially crafted compressed audio file to cause a denial of service.

Only deployments serving an audio-capable model through the chat endpoint are affected, and inline data URLs are not bounded by the audio fetch timeout.


Remediation

Install update from vendor's website.