CVE-2026-48746
vLLM is an inference and serving engine for large language models (LLMs). From 0.3.0 until 0.22.0, a vulnerability in ASGI web servers and s
CVSS
9.1
Critical
EPSS
1.2%
p65
KEV
—
Exploit Today
20
0-100
Published: Jun 22, 2026 · Last modified: Sep 1, 2026 · CWE-444 · CWE-501
1.2%EPSS · 30 days1.2%
2026-08-132026-09-10
vLLM is an inference and serving engine for large language models (LLMs). From 0.3.0 until 0.22.0, a vulnerability in ASGI web servers and starlette's trust on those web servers enables an authentication bypass of the OpenAI API AuthenticationMiddleware. It allows to use the API without providing the configured VLLM_API_KEY or --api-key. This vulnerability is fixed in 0.22.0.
- github.comhttps://github.com/vllm-project/vllm/pull/43426
- github.comhttps://github.com/vllm-project/vllm/security/advisories/GHSA-94f4-hr76-p5j6
- x41-dsec.dehttps://x41-dsec.de/lab/advisories/x41-2026-002-starlette
- access.redhat.comhttps://access.redhat.com/errata/RHSA-2026:30088
- access.redhat.comhttps://access.redhat.com/errata/RHSA-2026:30089
- access.redhat.comhttps://access.redhat.com/errata/RHSA-2026:36005
- access.redhat.comhttps://access.redhat.com/errata/RHSA-2026:36006
- access.redhat.comhttps://access.redhat.com/errata/RHSA-2026:42132
- access.redhat.comhttps://access.redhat.com/errata/RHSA-2026:42142
- access.redhat.comhttps://access.redhat.com/errata/RHSA-2026:42644
- access.redhat.comhttps://access.redhat.com/errata/RHSA-2026:43038
- access.redhat.comhttps://access.redhat.com/errata/RHSA-2026:61627
- access.redhat.comhttps://access.redhat.com/errata/RHSA-2026:61629
- access.redhat.comhttps://access.redhat.com/security/cve/CVE-2026-48746
- bugzilla.redhat.comhttps://bugzilla.redhat.com/show_bug.cgi?id=2491581
- security.access.redhat.comhttps://security.access.redhat.com/data/csaf/v2/vex/2026/cve-2026-48746.json
- x41-dsec.dehttps://x41-dsec.de/lab/advisories/x41-2026-002-starlette
CVECVSSEPSSKEVRExploitTitleMod.
CVE-2026-890446.5 MED—
——0Netty versions 4.1.133.Final through 4.1.137.Final and 4.2.13.Final through 4.2.17.Final fail to properly validate the final transfer coding in the Transfer-Encoding header, allowing attackers to smuggle requests by using malformed encoding declarations. Attackers can split Transfer-Encoding headers across multiple lines or use values like 'chunked, xchunked' to bypass validation and decode messages as chunked when the final coding is not chunked, enabling request smuggling attacks.1dCVE-2026-88009——
——0Traefik is an open source HTTP reverse proxy and load balancer. Prior to 2.11.57, and 3.7.13, Traefik accepts a rootless HTTP/1 request target that Go stores in URL.Opaque while leaving URL.Path empty. The rewriteRequestBuilder path evaluates routing, path sanitization, forwardAuth, encodedCharacters, and access logging against a path normalized to / but forwards URL.Opaque verbatim to the backend, allowing cross-vhost routing bypass, path-scoped authorization bypass, and access-log evasion when the backend interprets the opaque target as a path. This issue is fixed in 2.11.57 and 3.7.13.1dCVE-2026-88008——
——0Traefik is an open source HTTP reverse proxy and load balancer. From 2.11.26 until 2.11.57 and 3.7.13, Traefik forwards a client-supplied Connection header requesting Upgrade, the Upgrade: h2c token, and HTTP2-Settings to a shared backend. If the backend accepts h2c and returns 101 Switching Protocols, Traefik enters a raw tunnel and no longer applies routers, BasicAuth, ForwardAuth, IPAllowList, RateLimit, access logging, metrics, or tracing to later HTTP/2 requests, allowing an unauthenticated request through an unprotected route to reach protected paths on the same backend. This issue is fixed in 2.11.57 and 3.7.13.1dCVE-2026-813568.2 HIG25.7%
——8Inconsistent interpretation of http requests ('http request/response smuggling') in Visual Studio Code allows an unauthorized attacker to bypass a security feature over a network.1dCVE-2026-19203—21.3%
——6A client may issue specially crafted HTTP/1.1 chunked requests to a Jetty server that cause Jetty and an intermediary proxy to interpret different request boundaries, potentially resulting in HTTP request smuggling.
This is caused by Jetty accepting a lone LF character as a terminator in parts of chunked request parsing. Depending on the Jetty version and configured HTTP compliance mode, this may occur in chunk extensions, chunk data termination, or trailer termination.3dCVE-2026-185403.7 LOW12.2%
——4undici's retry interceptor can append the body of a ranged retry response to bytes already delivered from an earlier partial response while still presenting the original response's status and headers. This happens when an upstream server delivers part of a body without a trustworthy resume checkpoint, for example a non-success response whose headers were already sent or a partial-content response with an unusable content range, then closes the connection and answers the resumed range request with more bytes. As a result the response body can be longer than the Content-Length that the application observes. An application that relays such a response to a downstream HTTP/1.1 peer without normalizing the framing can emit a body that exceeds the forwarded Content-Length, and the excess bytes can be interpreted as the start of a following response, which enables downstream response splitting or desynchronization. Exploitation requires an attacker-controlled upstream server and an application that forwards the response through a framing-sensitive path. This affects undici versions before 6.28.1, from 7.0.0 up to 7.29.1, and from 8.0.0 up to 8.10.2. Users should upgrade to undici 6.28.1, 7.29.1, or 8.10.2.3d