In vllm versions >=0.7.0 and <0.19.0, the VideoMediaIO.load_base64() method improperly handles video/jpeg base64 data URLs by splitting them on commas to extract JPEG frames without enforcing the num_frames limit parameter. This bypass allows an attacker to supply a single API request with thousands of base64-encoded JPEG frames, causing the server to decode all frames into memory, leading to an out-of-memory crash. The vulnerability is classified under CWE-770 (Allocation of Resources Without Limits or Throttling). It has a CVSS 3.1 base score of 6.5 (medium severity) with network attack vector, low attack complexity, low privileges required, no user interaction, and impacts availability only. The vulnerability is fixed in vllm version 0.19.0. Red Hat references this CVE in their advisories but does not explicitly state a patch for vllm itself; users should upgrade to vllm 0.19.0 or later to remediate.

An attacker with the ability to send API requests to the vulnerable vllm server can cause a denial of service by exhausting server memory through a specially crafted request containing thousands of base64-encoded JPEG frames. This results in an out-of-memory crash, impacting availability. There is no impact on confidentiality or integrity according to the CVSS vector and description.

A fix is available by upgrading to vllm version 0.19.0 or later, where the frame count limit is properly enforced in the load_base64() method. Since the vendor advisory does not indicate any temporary fixes or workarounds, users should apply this upgrade to remediate the vulnerability. Patch status is confirmed by the vendor advisory and CVE description.