vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs.
CVE-2026-54234 is a high-severity improper input validation vulnerability in vllm from Vllm-project, with a CVSS 3.1 base score of 7.5 out of 10. It can be exploited remotely over the network by an unauthenticated attacker, without any action from a user. Affected vllm versions: < 0.24.0. The record does not name a fixed version, so check the vendor advisory. There is no confirmed exploitation so far. EPSS estimates a 0.62% chance of exploitation within 30 days, higher than 50% of scored CVEs.
High severity with no evidence of exploitation so far.
What to do
- Check which version of vllm you run against the affected versions below, and install the vendor's latest release.
- Read the advisory on github.com for workarounds and exact fixed builds.
- Until you can update, restrict network access to the affected service so it is not reachable from untrusted networks.
Affected products and versions
| Vendor | Product | Affected versions | Fixed in |
|---|---|---|---|
| Vllm-project | vllm | < 0.24.0 | Not stated |
Version data comes from the organisation that assigned the CVE. Confirm against the vendor advisory before deciding you are unaffected.
How the severity score breaks down
- Attack vector
- NetworkCan be exploited remotely over a network, including the internet.
- Attack complexity
- LowNo special conditions are needed; the attack works reliably.
- Privileges required
- NoneNo account or login is needed.
- User interaction
- NoneNo action by a victim is required.
- Scope
- UnchangedImpact stays within the vulnerable component.
- Confidentiality
- NoneNo data is disclosed.
- Integrity
- NoneNothing can be modified.
- Availability
- HighThe component can be taken down completely.
Vector: CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H. Scored by CNA (github.com).
Official description
vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs. Prior to 0.24.0, a frontend-legal multi-request speculative decoding workload can cause the rejection sampler to produce a recovered token equal to the model vocabulary size boundary value, which is then converted to negative one when the engine selects the next live token for a request and is written back into the drafter's input ids; that out-of-vocabulary value is later consumed by the model's embedding and attention path and crashes the engine worker with a GPU device-side assertion. The same triggering request sequence is reachable through the public gRPC Generate and Abort endpoints, so a remote client that can send generation requests can crash the shared engine worker, aborting concurrent requests and causing a service-wide denial of service for other clients of the deployment until the worker is restarted. This issue is fixed in version 0.24.0.
Source: the CVE record as published by its numbering authority, via NVD and CVE.org.
The underlying weakness: Improper Input Validation
The product receives input or data, but it does not validate or incorrectly validates that the input has the properties that are required to process the data safely and correctly.
How developers prevent this class of flaw
- Consider using language-theoretic security (LangSec) techniques that characterize inputs using a formal language and build recognizers for that language. This effectively requires parsing to be a distinct layer that effectively enforces a boundary between raw input and internal data representations, instead of allowing parser code to be scattered throughout the program, where it could be subject to errors or inconsistencies that create weaknesses.
- Use an input validation framework such as Struts or the OWASP ESAPI Validation API. Note that using a framework does not automatically address all input validation problems; be mindful of weaknesses that could arise from misusing the framework itself (CWE-1173).
- Understand all the potential areas where untrusted inputs can enter the product, including but not limited to: 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.
Timeline
- CVE record published.
- CISA published its SSVC assessment (exploitation: poc, automatable: yes, technical impact: partial).
- Record last changed materially (score, affected products, status or references).
References
Advisories
Patches and release notes
Other vulnerabilities in vllm
- 7.5 High severity
- 9.8 Critical severity
- 8.8 High severity
- 7.7 High severity
- 7.5 High severity
- 7.5 High severity
Recent Improper Input Validation vulnerabilities
- 10.0 Critical severity
- 7.3 High severity
- 7.3 High severity
- 7.3 High severity
- 9.8 Critical severity
- 7.5 High severity
Spotted an error? Tell us.