A path traversal flaw exists in deepset‑AI FARM’s EmbeddingTokenizer.from_pretrained function due to insufficient validation of the vocab_filename argument. By supplying a crafted filename that contains directory traversal sequences, an attacker can cause the tokenizer loader to resolve a path outside the intended directory, thereby reading arbitrary files on the host. If the downloaded vocabulary contains executable code, the attacker could also place files that could be later executed by other processes. The vulnerability permits remote exploitation because the function can be invoked over exposed interfaces in applications that expose FARM as a service.

The flaw affects deepset‑AI FARM versions up to and including 0.8.0. All releases in this series that are no longer supported by the maintainer are impacted. No supported (up‑to‑date) releases are known to contain the issue.

The CVSS score of 5.3 indicates a moderate severity. The EPSS score is not available, and the vulnerability is not listed in the CISA KEV catalog. Based on the description, the likely attack vector is remote. A published exploit demonstrates that the path traversal can be triggered from an attacker‐controlled environment, enabling the attacker to access arbitrary files on the system. Because the flaw resides in the library code, exploitation does not require additional privilege escalation beyond the permissions of the running process, making it readily exploitable for attackers who can reach the service.