CVE WatchOct 8, 2026

CVE-2026-105758

vLLM is an inference and serving engine for large language models. From 0.24.0 until 0.30.0, the Qwen2VLVideoBackend and Qwen3VLVideoBackend

CVSS

5.3

Medium

EPSS

0.3%

p22

KEV

—

Exploit Today

0-100

Published: Oct 5, 2026 · Last modified: Oct 8, 2026 · CWE-770

EPSS · 30d

0.3%EPSS · 30 days0.3%

2026-10-062026-10-21

Technical description

vLLM is an inference and serving engine for large language models. From 0.24.0 until 0.30.0, the Qwen2VLVideoBackend and Qwen3VLVideoBackend classes accept request-level values for the media_io_kwargs.video.max_frames and media_io_kwargs.video.fps fields without enforcing server-side ceilings. An unauthenticated caller can submit these values to the /tokenize endpoint, causing the sampler to decode every frame selected from attacker-controlled video input, consume disproportionate frontend memory, and potentially terminate the API process before scheduling or admission control. The Rust frontend is not affected because it rejects the media_io_kwargs field. This issue is fixed in version 0.30.0.

Official references

Related CVEs

CVECVSSEPSSKEVRExploitTitleMod.

CVE-2026-86696—

—

———Allocation of Resources Without Limits or Throttling vulnerability in esaml in dropbox esaml allows an unauthenticated attacker to exhaust the Erlang VM atom table and crash the node via unsigned LogoutResponse messages. esaml:status_code_map/1 in src/esaml.erl maps any StatusCode value it does not recognize to an atom with list_to_atom/1 on the last colon-separated segment of the URN. esaml_sp:validate_logout_response/2 verifies a signature only when the message carries one, so an unsigned LogoutResponse posted to the single-logout endpoint reaches esaml:decode_logout_response/1 and mints one permanent atom per distinct status value. Atoms are never reclaimed and the VM caps the atom table at about one million entries, so an attacker who posts that many distinct values aborts the node. This issue affects esaml: from 0.1.0 onward.17h CVE-2026-105270—

—

———Allocation of Resources Without Limits or Throttling vulnerability in esaml_binding in dropbox esaml allows an unauthenticated attacker to exhaust the Erlang VM atom table and crash the node via a crafted SAMLResponse. esaml_binding:decode_response/2 parses the decoded SAMLResponse with xmerl_scan:string/2, which interns every element name, attribute name, and namespace prefix it reads as a permanent atom. The SAML consume (ACS) and single-logout endpoints call it on unauthenticated requests before any signature check, so one well-formed document with thousands of distinct names mints thousands of atoms while staying inside the 128 KB request body limit. Atoms are never reclaimed and the VM caps the atom table at about one million entries, so a few dozen requests abort the node. This issue affects esaml: from 0.1.0 onward.19h CVE-2026-1147315.3 MED

—

———Roundcube Webmail before 1.6.20 and 1.7.x before 1.7.5 is affected by an uncontrolled resource consumption vulnerability in the CSS proxy (the utils/modcss action). The fetched stylesheet was read into memory in full, with no size limit and no read timeout, so a stylesheet URL referencing an attacker-controlled host could be used to exhaust server memory. Upstream fixed this in the commit “Add a size limit to the CSS proxy content processing”, which caps the response at 5 MB, reads it as a stream and applies a read timeout.23h CVE-2026-963775.3 MED

—

———Allocation of Resources Without Limits or Throttling vulnerability in Drupal Combined image style allows Excessive Allocation. This issue affects Combined image style versions: from 0.0.0 to 1.0.7.11h CVE-2026-963657.5 HIG

—

———Allocation of Resources Without Limits or Throttling vulnerability in Drupal Webform allows Excessive Allocation. This issue affects Webform versions: from 0.0.0 to 6.2.12, from 6.3.0 to 6.3.1.11h CVE-2026-1147233.7 LOW

—

———Roundcube Webmail before 1.6.20 and 1.7.x before 1.7.5 allows the ImageMagick resource limits to be bypassed during thumbnail generation. When rcube_image falls back to GD, the available-memory check was applied only on Windows, so on other platforms a crafted image with large dimensions was processed by GD without any resource limit and could exhaust server memory. Upstream fixed this in the commit “Fix ImageMagick resource limits bypass by GD fallback during thumbnail generation”, which applies the memory check on all platforms.1d