GHSA-h3rc-6mm3-gc2m CVE-2026-114053

vLLM: Request-controlled mm_processor_kwargs.code_revision allows remote code execution

Critical PyPI

Severity: 9.2 / 10; Critical · CVSS 4.0
Exploit chance (EPSS): 0.46%; Chance of exploitation in the next 30 days. Higher than 40.04% of scored vulnerabilities.
Known exploited: No; Not in CISA's catalog
Weakness: CWE-829 Inclusion of Functionality from Untrusted Control Sphere; CWE-94 Improper Control of Generation of Code ('Code Injection')
Fix: 0.31.0; Upgrade vllm to this version or later

Description

Summary

An untrusted multimodal request can set mm_processor_kwargs.code_revision. When vLLM serves a model whose Hugging Face processor uses dynamic auto_map code and the operator enabled --trust-remote-code, this request value reaches AutoProcessor.from_pretrained and selects the Python revision imported by Transformers.

This lets a client replace the operator-reviewed processor module with code from another revision of the same model repository. The operator-selected model/data revision remains pinned; only the executed processor code changes.

Importing a Python module is arbitrary code execution in the vLLM API/renderer process. The PoC below demonstrates this with an import-time filesystem write. It is fully offline and does not require a GPU, model weights, or changes to a real Hub repository.

Affected snapshot

Repository: vllm-project/vllm
Verified commit: 570735520942b6428ae98f1eb442c6d88adae55e
Verified with Transformers 5.15.0 and 5.16.1
Affected release range has not been bisected

Affected configuration

The vulnerable path exists when all of the following are true:

The served model's processor uses dynamic auto_map.AutoProcessor code.
The operator enabled trust_remote_code.
The server can fetch the request-selected revision, or it is already cached.
The attacker can create or otherwise control a revision in the served model repository. An example is a Hugging Face community pull-request ref when that repository permits them.
A multimodal request reaches the processor path.

The default trust_remote_code=False configuration is not affected.

Attack surface

mm_processor_kwargs is accepted directly by public request models:

/tokenize is particularly relevant because vLLM's authentication middleware guards only /v1, /v2, /inference, and /cohere prefixes:

AuthenticationMiddleware

Consequently, /tokenize remains unauthenticated even when --api-key is configured, unless an external proxy protects it. The vLLM Security Guide documents /tokenize as an unprotected endpoint.

Root cause

The request value follows this path:

OnlineRenderer.preprocess_chat places request mm_processor_kwargs in prompt extras.
Multimodal rendering forwards those kwargs into processor construction through MMProcessorInputs and _process_multimodal_async.
MultiModalConfig.merge_mm_processor_kwargs overlays inference values on startup values, so the request wins.
cached_get_processor_without_dynamic_kwargs filters only recognized processor-call keys. code_revision is retained as a processor-construction kwarg.
cached_processor_from_config pins revision and trust_remote_code, but forwards the remaining kwargs without reserving code_revision.
get_processor calls AutoProcessor.from_pretrained(..., revision=revision, trust_remote_code=trust_remote_code, **kwargs).
Transformers AutoProcessor.from_pretrained uses code_revision to retrieve and import the dynamic module through get_class_from_dynamic_module, then removes it before loading processor data. This is why code comes from the attacker revision while configuration/data remain operator-pinned.

Request-time processor arguments are not separated from processor-construction arguments, and code_revision is not reserved.

Suggested remediation

Separate processor-construction arguments from untrusted processor-call arguments. An allowlist of supported request-time preprocessing controls is safest.
Reject code_revision and other loader/source controls in request mm_processor_kwargs, and pass model_config.code_revision explicitly as the only permitted processor-code revision.
Add a regression test proving a request cannot change the loaded module revision while operator pins remain effective.
Protect /tokenize with the configured authentication policy as defense in depth.

Expected safe behavior is either a 4xx response for reserved request keys or silently ignoring them before processor construction. Rejection is preferable because it makes attempted misuse visible.

A patch for this was merged here: https://github.com/vllm-project/vllm/pull/58830

Affected versions

PyPI

Package	Vulnerable	Fixed in
vllm	`>= 0.7.3, < 0.31.0`	`0.31.0`

How to fix

Move to the fixed version or later with your package manager, update the lockfile, and run your tests. If the package only arrives through another dependency, upgrade that dependency first. If none of its releases pulls a fixed version yet, force the version with your package manager's override.

vllm PyPI → `0.31.0`

pip

pip install --upgrade "vllm>=0.31.0"

poetry

poetry add "vllm>=0.31.0"

uv

uv add "vllm>=0.31.0"

Severity breakdown

CVSS 4.0 9.2

CVSS:4.0/AV:N/AC:L/AT:P/PR:N/UI:N/VC:H/VI:H/VA:H/SC:N/SI:N/SA:N/E:X/CR:X/IR:X/AR:X/MAV:X/MAC:X/MAT:X/MPR:X/MUI:X/MVC:X/MVI:X/MVA:X/MSC:X/MSI:X/MSA:X/S:X/AU:X/R:X/V:X/RE:X/U:X

Attack vector: Network
Attack complexity: Low
Attack requirements: Present
Privileges required: None
User interaction: None
Confidentiality (vulnerable system): High
Integrity (vulnerable system): High
Availability (vulnerable system): High
Confidentiality (subsequent systems): None
Integrity (subsequent systems): None
Availability (subsequent systems): None
Exploit maturity: Not defined
Scope: X

CVSS 3.1 8.1

CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:H/I:H/A:H

Attack vector: Network
Attack complexity: High
Privileges required: None
User interaction: None
Scope: Unchanged
Confidentiality: High
Integrity: High
Availability: High

Weakness

CWE-829: Inclusion of Functionality from Untrusted Control Sphere

The product imports, requires, or includes executable functionality (such as a library) from a source that is outside of the intended control sphere.

How to prevent it

Architecture and Design: Use a vetted library or framework that does not allow this weakness to occur or provides constructs that make this weakness easier to avoid [REF-1482].
Architecture and Design: When the set of acceptable objects, such as filenames or URLs, is limited or known, create a mapping from a set of fixed input values (such as numeric IDs) to the actual filenames or URLs, and reject all other inputs. For example, ID 1 could map to "inbox.txt" and ID 2 could map to "profile.txt". Features such as the ESAPI AccessReferenceMap [REF-45] provide this capability.
Architecture and Design: For any security checks that are performed on the client side, ensure that these checks are duplicated on the server side, in order to avoid CWE-602. Attackers can bypass the client-side checks by modifying values after the checks have been performed, or by changing the client to remove the client-side checks entirely. Then, these modified values would be submitted to the server.
Architecture and Design, Operation: Run the code in a "jail" or similar sandbox environment that enforces strict boundaries between the process and the operating system. This may effectively restrict which files can be accessed in a particular directory or which commands can be executed by the software. OS-level examples include the Unix chroot jail, AppArmor, and SELinux. In general, managed code may provide some protection. For example, java.io.FilePermission in the Java SecurityManager allows the software to specify restrictions on file operations. This may not be a feasible solution, and it only limits the impact to the operating system; the rest of the application may still be subject to compromise. Be careful to avoid CWE-243 and other weaknesses related to jails.
Architecture and Design, Operation: Run your code using the lowest privileges that are required to accomplish the necessary tasks [REF-76]. If possible, create isolated accounts with limited privileges that are only used for a single task. That way, a successful attack will not immediately give the attacker access to the rest of the software or its environment. For example, database applications rarely need to run as the database administrator, especially in day-to-day operations.
Implementation: Assume all input is malicious. Use an "accept known good" input validation strategy, i.e., use a list of acceptable inputs that strictly conform to specifications. Reject any input that does not strictly conform to specifications, or transform it into something that does. When performing input validation, consider all potentially relevant properties, including length, type of input, the full range of acceptable values, missing or extra inputs, syntax, consistency across related fields, and conformance to business rules. As an example of business rule logic, "boat" may be syntactically valid because it only contains alphanumeric characters, but it is not valid if the input is only expected to contain colors such as "red" or "blue." Do not rely exclusively on looking for malicious or malformed inputs. This is likely to miss at least one undesirable input, especially if the code's environment changes. This can give attackers enough room to bypass the intended validation. However, denylists can be useful for detecting potential attacks or determining which inputs are so malformed that they should be rejected outright. When validating filenames, use stringent allowlists that limit the character set to be used. If feasible, only allow a single "." character in the filename to avoid weaknesses such as CWE-23, and exclude directory separators such as "/" to avoid CWE-36. Use a list of allowable file extensions, which will help to avoid CWE-434. Do not rely exclusively on a filtering mechanism that removes potentially dangerous characters. This is equivalent to a denylist, which may be incomplete (CWE-184). For example, filtering "/" is insufficient protection if the filesystem also supports the use of "\" as a directory separator. Another possible error could occur when the filtering is applied in a way that still produces dangerous data (CWE-182). For example, if "../" sequences are removed from the ".../...//" string in a sequential fashion, two instances of "../" would b
Architecture and Design, Operation: Store library, include, and utility files outside of the web document root, if possible. Otherwise, store them in a separate directory and use the web server's access control capabilities to prevent attackers from directly requesting them. One common practice is to define a fixed constant in each calling program, then check for the existence of the constant in the library/include file; if the constant does not exist, then the file was directly requested, and it can exit immediately. This significantly reduces the chance of an attacker being able to bypass any protection mechanisms that are in the base program but not in the include files. It will also reduce the attack surface.
Architecture and Design, Implementation: Understand all the potential areas where untrusted inputs can enter your software: parameters or arguments, cookies, anything read from the network, environment variables, reverse DNS lookups, query results, request headers, URL components, e-mail, files, filenames, databases, and any external systems that provide data to the application. Remember that such inputs may be obtained indirectly through API calls. Many file inclusion problems occur because the programmer assumed that certain inputs could not be modified, especially for cookies and URL components.
Operation: Use an application firewall that can detect attacks against this weakness. It can be beneficial in cases in which the code cannot be fixed (because it is controlled by a third party), as an emergency prevention measure while more comprehensive software assurance measures are applied, or to provide defense in depth [REF-1481].

CWE-94: Improper Control of Generation of Code ('Code Injection')

The product constructs all or part of a code segment using externally-influenced input from an upstream component, but it does not neutralize or incorrectly neutralizes special elements that could modify the syntax or behavior of the intended code segment.

How to prevent it

Architecture and Design: Refactor your program so that you do not have to dynamically generate code.
Architecture and Design: Run your code in a "jail" or similar sandbox environment that enforces strict boundaries between the process and the operating system. This may effectively restrict which code can be executed by your product. Examples include the Unix chroot jail and AppArmor. In general, managed code may provide some protection. This may not be a feasible solution, and it only limits the impact to the operating system; the rest of your application may still be subject to compromise. Be careful to avoid CWE-243 and other weaknesses related to jails.
Implementation: Assume all input is malicious. Use an "accept known good" input validation strategy, i.e., use a list of acceptable inputs that strictly conform to specifications. Reject any input that does not strictly conform to specifications, or transform it into something that does. When performing input validation, consider all potentially relevant properties, including length, type of input, the full range of acceptable values, missing or extra inputs, syntax, consistency across related fields, and conformance to business rules. As an example of business rule logic, "boat" may be syntactically valid because it only contains alphanumeric characters, but it is not valid if the input is only expected to contain colors such as "red" or "blue." Do not rely exclusively on looking for malicious or malformed inputs. This is likely to miss at least one undesirable input, especially if the code's environment changes. This can give attackers enough room to bypass the intended validation. However, denylists can be useful for detecting potential attacks or determining which inputs are so malformed that they should be rejected outright. To reduce the likelihood of code injection, use stringent allowlists that limit which constructs are allowed. If you are dynamically constructing code that invokes a function, then verifying that the input is alphanumeric might be insufficient. An attacker might still be able to reference a dangerous function that you did not intend to allow, such as system(), exec(), or exit().
Testing: Use dynamic tools and techniques that interact with the product using large test suites with many diverse inputs, such as fuzz testing (fuzzing), robustness testing, and fault injection. The product's operation may slow down, but it should not become unstable, crash, or generate incorrect results.
Operation: Run the code in an environment that performs automatic taint propagation and prevents any command execution that uses tainted variables, such as Perl's "-T" switch. This will force the program to perform validation steps that remove the taint, although you must be careful to correctly validate your inputs so that you do not accidentally mark dangerous inputs as untainted (see CWE-183 and CWE-184).
Operation: Run the code in an environment that performs automatic taint propagation and prevents any command execution that uses tainted variables, such as Perl's "-T" switch. This will force the program to perform validation steps that remove the taint, although you must be careful to correctly validate your inputs so that you do not accidentally mark dangerous inputs as untainted (see CWE-183 and CWE-184).
Implementation: For Python programs, it is frequently encouraged to use the ast.literal_eval() function instead of eval, since it is intentionally designed to avoid executing code. However, an adversary could still cause excessive memory or stack consumption via deeply nested structures [REF-1372], so the python documentation discourages use of ast.literal_eval() on untrusted data [REF-1373].

References

Advisories

nvd.nist.gov/vuln/detail/cve-2026-114053

Packages

github.com/vllm-project/vllm

Timeline

Oct 21, 2026 Published
Oct 27, 2026 Last updated by the source
Oct 31, 2026 Latest EPSS score

Advisory	Severity	Why it's related
GHSA-8397-8qf6-f2gm vLLM: Phi-4 multimodal audio encoder materializes and retains full attention matrices: a single max-length audio OOM-crashes the vLLM engine	Medium	Also affects vllm
GHSA-x567-49g5-p4cf vLLM: Responses API bypasses trust_request_chat_template for chat_template_kwargs	Medium	Also affects vllm
GHSA-vvpg-rrww-wpmh vLLM: Permissive default CORS origins allow cross-origin requests to the vLLM HTTP server	Medium	Also affects vllm
GHSA-q97c-p368-75c4 vLLM: Out-of-vocabulary prompt token id passes input validation and triggers a CUDA device-side assert (engine crash / DoS)	Medium	Also affects vllm
GHSA-h3rc-j82p-2c3h vLLM: Untrusted EPS inputs can trigger denial of service or Ghostscript RCE	High	Also affects vllm
GHSA-gq5q-89c7-vhvx vLLM DeepSeek-OCR image processing vulnerable to denial of service	Medium	Also affects vllm
GHSA-6757-hggm-hp2w vLLM: Public multimodal SHM handle forgery reaches worker-side `pickle.loads()` with insecure serialization disabled	High	Also affects vllm
GHSA-542m-9234-8gvf vLLM: DoS via unbounded chat_template Jinja2 rendering exhausts the renderer thread pool	Medium	Also affects vllm

Is this in your code?

MergeHarbor checks every repository's dependencies against this database, ranks what matters by exploitability and reachability, routes it to the owning team and opens the fix pull request.

Check your repos with MergeHarbor