The vulnerability resides in the Hugging Face Text Embeddings Inference service up to version 1.9.4. An unchecked argument in the router’s Args struct causes the api_key to be written to plain‑text logs. When an attacker can read router logs, container stdout, or OTLP telemetry they can extract the Bearer token, then use it to call protected embedding and reranking endpoints without authorization. This results in credential disclosure and potential abuse.

The affected product is Hugging Face Text Embeddings Inference, with the issue present through version 1.9.4. Any deployment that uses an earlier or equal version inherits the flaw, as the router code does not apply a redact attribute to the api_key field. The vulnerability applies to every instance that logs requests.

The CVSS score of 4.8 indicates a low‑to‑moderate severity, and no EPSS data or KEV listing are available. The potential impact depends on attacker ability to access logs, which is feasible in compromised or mis‑configured environments. The flaw does not allow arbitrary code execution but exposes authentication credentials that grant access to all protected endpoints, making the risk significant in high‑value deployments.