vLLM Remote DoS via Uncontrolled Resource Consumption in Multimodal Processing

vLLM, a popular library for LLM inference, contains a high-severity vulnerability classified as uncontrolled resource consumption. The flaw impacts versions starting from 0.7.3 and ending before 0.31.0, exposing shared inference servers to denial-of-service attacks.

What happened

According to the National Vulnerability Database record for CVE-2026-114055, the issue stems from inadequate validation of the mm_processor_kwargs parameter. This parameter allows remote clients to supply configuration details for multimodal processors within the vLLM framework. The lack of restriction means attackers can inject arbitrary values into these settings.

Specifically, adversaries can set excessively large values for image size, patch count, crop count, or frame count. These manipulations target processors such as Isaac, InternVL, and H2OVL. The resulting computational load is designed to exhaust available system memory.

Why it matters

Shared inference servers are particularly vulnerable because multiple clients often share the same computational resources. When one client forces the system to process an abnormally large workload, it consumes memory that other legitimate users require. This leads to service degradation or complete termination of the shared server.

The CVSS score of 8.7 indicates a high severity rating. The weakness aligns with CWE-400, which covers uncontrolled resource consumption. This type of flaw is critical in multi-tenant environments where isolation between clients is not guaranteed by the application layer.

Next steps

• Identify all instances of vLLM running versions prior to 0.31.0.
• Review configuration files for exposed mm_processor_kwargs parameters.
• Monitor shared servers for unusual spikes in memory usage.
• Isolate affected systems from untrusted network segments until patched.

Detection notes

Teams usually detect this activity through sudden, unexplained increases in memory consumption on inference servers. Check whether application logs show repeated requests with unusually large parameter values for image or video processing. Look for alerts indicating out-of-memory errors or process terminations. Monitor network traffic for requests originating from unexpected sources.

Verify if specific multimodal processors like Isaac or InternVL are handling disproportionate loads. Contain potential incidents by temporarily restricting access to the affected endpoints. Analyze request patterns to identify the source of the malicious payloads. Ensure logging captures the full content of mm_processor_kwargs arguments for forensic review. Investigate any client accounts that deviate significantly from normal usage baselines.

What to do and how to stay safe: vLLM

• Audit current vLLM installations to confirm version numbers against the affected range.
• Implement strict input validation for all multimodal processing parameters.
• Set resource limits on individual requests to prevent single-tenant exhaustion.
• Monitor system memory usage closely for anomalies during peak processing times.

Step-by-step guide: Emergency Patching: How to Fix Critical Flaws Without Breaking Production

General security guidance from the Packet Tribune newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

Which versions of vLLM are affected by CVE-2026-114055?

Versions from 0.7.3 before 0.31.0 are affected.

What is the CVSS score for this vulnerability?

The CVSS score is 8.7, rated as HIGH.

How can attackers exploit this flaw?

Attackers can supply large values for image size, patch count, crop count, or frame count to exhaust memory.