CVE-2026-24779: vLLM vulnerable to Server-Side Request Forgery (SSRF) through MediaConnector
vLLM before 0.14.1 can be tricked into making requests to internal hosts. MediaConnector checks a user-supplied media URL's host with one parser and fetches it with another, and the two read backslashes differently, so the allowed-domain check can be bypassed. A signed-in API user can use this against internal services. Fixed in 0.14.1.
Published Jan 28, 2026 · Updated Jul 21, 2026 · Analysis Nov 1, 2026
What an attacker can do
MediaConnector.load_from_url and load_from_url_async fetch images, audio and video from URLs in API requests. They validate the host with urlparse from Python's urllib (RFC 3986), in _assert_url_in_allowed_media_domains. The fetch then goes through requests, which uses urllib3's parse_url (WHATWG-style). The two disagree about backslashes in a URL, so a URL can look like an allowed host to the check and resolve to a different host in the request.
The attacker needs access to the vLLM API with low privileges and a multimodal request that carries a URL. The vLLM server then sends requests to hosts the allowlist was meant to keep off limits: internal services, metadata endpoints and other pods. The CVSS vector rates confidentiality high and availability low.
The advisory singles out containerized setups like llm-d. A compromised vLLM pod could scan the internal network, talk to other pods, or send false metrics such as KV cache state to an internal management endpoint, causing instability. The restriction only matters if you configured allowed media domains; if you did not, the fetch is open anyway.
Who is affected
vllm below 0.14.1, on deployments that accept media URLs in multimodal requests and rely on the media domain allowlist to limit where the server can fetch from. Kubernetes and llm-d style clusters carry the most risk, because internal services are reachable from the pod. Check the version:
pip show vllmFor containers, read the image tag. Red Hat has published errata for its builds.
vllm
- Ecosystem
- PyPI
- Affected
- < 0.14.1
- Fixed in
- 0.14.1
- Check
pip show vllm
How to fix it
Upgrade to vllm 0.14.1.
pip install --upgrade "vllm>=0.14.1"The fix, in pull request 32746, replaces urlparse with urllib3's parse_url in vllm/multimodal/utils.py, so the check and the request use the same parser. The advisory gives no other workaround. Until you upgrade, add network egress rules so the vLLM pod cannot reach internal services or metadata addresses, and do not expose the API to untrusted callers.
How to tell if you were hit
Look in API request logs for media URLs containing backslashes, or hosts that do not match your allowed domains. Check egress and firewall logs from the vLLM host for connections to internal ranges or cloud metadata addresses that it does not normally contact. In Kubernetes, review network policy denials and unexpected traffic from vLLM pods to management endpoints. The advisory gives no log signature.
Original advisory
From the GitHub Advisory Database (CC BY 4.0).
Summary
A Server-Side Request Forgery (SSRF) vulnerability exists in the MediaConnector class within the vLLM project's multimodal feature set. The load_from_url and load_from_url_async methods obtain and process media from URLs provided by users, using different Python parsing libraries when restricting the target host. These two parsing libraries have different interpretations of backslashes, which allows the host name restriction to be bypassed. This allows an attacker to coerce the vLLM server into making arbitrary requests to internal network resources.
This vulnerability is particularly critical in containerized environments like llm-d, where a compromised vLLM pod could be used to scan the internal network, interact with other pods, and potentially cause Denial of Service or access sensitive data. For example, an attacker could make the vLLM pod send malicious requests to an internal llm-d management endpoint, leading to system instability by falsely reporting metrics like the KV cache state.
Details
The core of the vulnerability lies in the MediaConnector.load_from_url method and its asynchronous counterpart. These methods accept a URL string to fetch media content (images, audio, video).
def load_from_url( > self, > url: str, > media_io: MediaIO[_M], > *, > fetch_timeout: int | None = None, > ) -> _M: # type: ignore[type-var] > url_spec = urlparse(url) > > if url_spec.scheme.startswith("http"): > self._assert_url_in_allowed_media_domains(url_spec) > > connection = self.connection > data = connection.get_bytes( > url, > timeout=fetch_timeout, > allow_redirects=envs.VLLM_MEDIA_URL_ALLOW_REDIRECTS, > ) > > return media_io.load_bytes(data)
The URL validation uses the urlparse function from Python's urllib module, while the request is made using the request function from Python's requests module. The requests module's underlying URL parsing is implemented using the parse_url function from Python's urllib3. These two parsing functions follow different URL specifications; one is implemented according to the RFC 3986 specification, and the other is implemented according to the WHATWG Living Standard. There is a difference in how the two functions handle backslashes (\) in URLs, which allows the hostname restriction to be bypassed.
Fix
- https://github.com/vllm-project/vllm/pull/32746
Credited to leishilong, leung-yao, Isotr0py, russellb.
Questions
- What is the root cause of CVE-2026-24779?
Two URL parsers disagree. The host check uses urllib's urlparse and the fetch uses requests with urllib3's parse_url. They read backslashes differently, so the check can be passed with a URL that goes elsewhere.
- Does it need authentication?
The CVSS vector requires low privileges, meaning access to the API. The advisory does not say more about authentication.
- Which vLLM version fixes it?
vllm 0.14.1 fixes it. Every earlier version is listed as vulnerable. Red Hat has also published errata for its own builds.
References
Read for this analysis:
- https://github.com/vllm-project/vllm/security/advisories/GHSA-qh4c-xf7m-gxfc
- https://github.com/vllm-project/vllm/pull/32746
- https://nvd.nist.gov/vuln/detail/CVE-2026-24779
- https://github.com/vllm-project/vllm/commit/f46d576c54fb8aeec5fc70560e850bed38ef17d7
- https://security.access.redhat.com/data/csaf/v2/vex/2026/cve-2026-24779.json
- https://pypi.org/project/vllm
- https://github.com/vllm-project/vllm
- https://github.com/pypa/advisory-database/tree/main/vulns/vllm/PYSEC-2026-2020.yaml
- https://github.com/advisories/GHSA-qh4c-xf7m-gxfc
- https://bugzilla.redhat.com/show_bug.cgi?id=2433624
- https://access.redhat.com/security/cve/CVE-2026-24779
- https://access.redhat.com/errata/RHSA-2026:3782
- https://access.redhat.com/errata/RHSA-2026:3462
- https://access.redhat.com/errata/RHSA-2026:3461
- https://access.redhat.com/errata/RHSA-2026:30089
- https://access.redhat.com/errata/RHSA-2026:30088
- https://access.redhat.com/errata/RHSA-2026:30087
- https://access.redhat.com/errata/RHSA-2026:19712
- https://access.redhat.com/errata/RHSA-2026:10184
- https://access.redhat.com/errata/RHSA-2026:42644
How this page is made: the data panel comes from the GitHub Advisory Database, CISA’s KEV catalog, the NVD and FIRST EPSS, and updates twice a day. The analysis is written by an AI model from the sources listed above and checked by automated rules before it is published. Check versions against the vendor's advisory before you act. Methodology
More vLLM vulnerabilities
All vLLM entriesCVE-2026-115911
vLLM: DoS via unbounded chat_template Jinja2 rendering exhausts the renderer thread pool
medium6.5Oct 28, 2026CVE-2026-115912
vLLM: Public multimodal SHM handle forgery reaches worker-side `pickle.loads()` with insecure serialization disabled
high7.5Oct 28, 2026CVE-2026-115913
vLLM: Phi-4 multimodal audio encoder materializes and retains full attention matrices: a single max-length audio OOM-crashes the vLLM engine
medium6.5Oct 28, 2026CVE-2026-115914
vLLM DeepSeek-OCR image processing vulnerable to denial of service
medium6.5Oct 28, 2026CVE-2026-115915
vLLM: Untrusted EPS inputs can trigger denial of service or Ghostscript RCE
high8.1Oct 28, 2026CVE-2026-115916
vLLM: Out-of-vocabulary prompt token id passes input validation and triggers a CUDA device-side assert (engine crash / DoS)
medium6.5Oct 28, 2026
