Technical impact
vLLM is an inference and serving engine that processes image inputs with Pillow. Prior to version 0.32.0, the ImageMediaIO component forwards raw image bytes to Pillow via Image.open(BytesIO(data)) without choosing a decoder, and the MIME type does not limit the selected format. When the image‑capable request paths (parse_image() and fetch_image()) receive an attacker‑controlled EPS payload, Pillow can interpret it as EPS and trigger Ghostscript on the host. If the host’s Ghostscript is vulnerable to CVE-2024-29510, the payload can be executed with the privileges of the vLLM process, yielding remote code execution; environments that use Pillow 12.0.0 through 12.2.0 can also trigger CVE-2026-59203, causing media‑loading worker exhaustion and a denial of service. Deployments that do not include the vulnerable Ghostscript or Pillow libraries are not affected by these specific impacts.
Affected Systems
This vulnerability applies to any installation of the vllm project running vLLM before 0.32.0. Impact is present only when the host has a Ghostscript version vulnerable to CVE-2024-29510, or when Pillow 12.0.0–12.2.0 is in use, allowing the execution of the EPS payload or worker exhaustion, respectively. Systems that do not rely on these vulnerable dependencies are not impacted.
Risk and Exploitability
The flaw carries a CVSS score of 8.1, indicating high severity. The EPSS score is less than 1%, showing a very low but non‑zero likelihood of exploitation. It is not listed in the CISA KEV catalog. A remote attacker who can send image data to an exposed vLLM inference endpoint can trigger the effect. If the host’s Ghostscript is vulnerable, the attacker can obtain execution privileges of the vLLM process; otherwise, the attacker may cause a denial of service by exhausting media‑loading workers. The likely attack vector is remote, as the description indicates exploitation occurs via API calls that accept image payloads from external sources.
Github GHSA