The vulnerability in vLLM (vllm-project/vllm) arises from the VideoMediaIO.load_base64() method's handling of video/jpeg data URLs. The method splits the base64 data string on commas to extract individual JPEG frames but does not enforce any limit on the number of frames processed. An attacker can exploit this by crafting a single API request containing thousands of comma-separated base64-encoded JPEG frames, causing the server to decode all frames into memory simultaneously. This results in uncontrolled resource consumption leading to an Out-of-Memory (OOM) Denial of Service (DoS) condition. The vulnerability is reachable via the OpenAI-compatible chat completions API and requires no authentication, increasing its risk. The CVSS v3.0 score is 7.5 (High), reflecting network attack vector, low attack complexity, no privileges or user interaction required, and an impact limited to availability (denial of service). No official patch or remediation level has been confirmed by the vendor. The Red Hat advisory linked does not provide explicit patch or mitigation details.

Successful exploitation results in a denial of service condition due to excessive memory consumption when processing maliciously crafted video/jpeg data URLs containing large numbers of base64-encoded JPEG frames. This can cause the server to crash or become unresponsive, impacting availability. There is no reported impact on confidentiality or integrity.

Patch status is not yet confirmed — check the vendor advisory for current remediation guidance. Since no official fix or workaround is documented in the vendor advisory, users should monitor the vendor's security advisories for updates. As a temporary measure, restricting or validating input sizes and frame counts in the VideoMediaIO.load_base64() method could reduce risk, but such mitigations require code changes and are not officially provided. Network-level protections such as rate limiting or filtering suspicious API requests may help reduce exposure until a patch is available.