CVE-2026-114055

What it actually does

Remote unauthenticated clients can supply arbitrary mm_processor_kwargs values (large image_size, patch_count, crop_count, or frame_count) to multimodal processors, causing unbounded memory allocation that leads to out-of-memory termination of the shared inference server.

Why it matters to the business

Attackers can repeatedly crash the shared AI inference service, resulting in denial of service, lost inference capacity, and operational disruption for any downstream applications relying on the server.

How an attack unfolds

An attacker sends crafted inference requests containing oversized mm_processor_kwargs parameters to vulnerable multimodal processors; the server allocates excessive memory per request until the process terminates or becomes unresponsive.

How to fix it

Upgrade vLLM to version 0.31.0 or later. Until then, enforce strict input validation and sanitization on all incoming mm_processor_kwargs fields at the request gateway or application layer, and apply resource limits or rate limiting per client.

How to confirm the fix worked

Confirm the installed vLLM version is 0.31.0 or newer. Test by submitting requests with deliberately large mm_processor_kwargs values and verify that the server rejects or safely caps the parameters instead of allocating unbounded memory.

Typical attack path

1. Discover an exposed vLLM inference endpoint accepting multimodal requests
2. Craft a request that sets extremely large values for image_size, patch_count, crop_count or frame_count inside mm_processor_kwargs
3. Submit one or more such requests to trigger progressive memory exhaustion
4. Observe server process termination or unresponsiveness confirming successful denial of service

How to detect it

Review inference request logs for entries containing mm_processor_kwargs with unusually large numeric values for image_size, patch_count, crop_count, or frame_count; correlate with memory-usage spikes or process restarts in system logs.

• mm_processor_kwargs.*image_size.*[0-9]{5,}
• mm_processor_kwargs.*patch_count.*[0-9]{5,}
• mm_processor_kwargs.*frame_count.*[0-9]{5,}

What your security team should do

• Immediately upgrade all vLLM deployments to 0.31.0 or later
• Add server-side validation and bounds checking for all multimodal processor parameters
• Implement per-client rate limiting and memory quotas on inference endpoints
• Monitor memory usage metrics and set alerts for sudden spikes correlated with request patterns