GROBID, a machine‑learning tool for extracting data from scholarly PDFs, contains an OS command injection flaw in its non‑server PDF conversion path. When a PDF file with a specially crafted filename is processed, the filename is interpolated into a bash command string. A single quote in the name can terminate the intended quoting, and the resulting command is executed using ProcessBuilder with the privileges of the GROBID process. The vulnerability allows an attacker to run arbitrary commands on the host, potentially leading to full system compromise.

All installations of grobidOrg:grobid with a version older than 0.9.1 using the Linux non‑server PDF conversion route are vulnerable. The bundled REST server path is not affected, but any use of the non‑server path with untrusted filenames exposes the system to exploitation.

The CVSS score is not disclosed, and the EPSS score is unavailable. The vulnerability is not listed in CISA's KEV catalog. Nonetheless, the flaw permits arbitrary command execution with GROBID process privileges, representing a high‑risk exposure. Attackers need only supply a PDF file with a malicious filename; the execution path is directly chained to the ProcessRunner without additional authentication or input checks. The attack vector is inferred to be local file handling, typically through a client uploading a PDF via any interface that triggers the non‑server conversion workflow.