If an application utilizing this library allows untrusted user input to influence the image source parameter, an attacker can supply local filesystem paths. When the library processes these requests, it treats the provided path as a file to be read by the application process. The contents of these files are then included in the payload sent to the configured NVIDIA/NIM model backend. This allows attackers to exfiltrate arbitrary files that the application process has permissions to read, such as configuration files, local credentials, or source code. The vulnerability was addressed in version 1.4.2, which strictly filters input types to only permit remote URLs, base64-encoded data URIs, and authorized asset identifiers.

Impact

Applications built with langchain-nvidia-ai-endpoints that process untrusted image inputs are at risk of arbitrary file disclosure. If exploited, an attacker can gain unauthorized access to sensitive local files, which may lead to further exploitation, credential theft, or exposure of internal infrastructure details. The potential impact depends on the filesystem permissions assigned to the service account running the application process.

Recommendation

• Upgrade langchain-nvidia-ai-endpoints to version 1.4.2 or later immediately.
• If patching is delayed, implement strict server-side validation to reject any user-supplied image input that resembles a local filesystem path (e.g., paths starting with /, C:\, or relative path traversal sequences like ../).
• Enforce the principle of least privilege for the application process; ensure the service account running the application has read access restricted only to the files and directories strictly necessary for its operation.
• Review application logs for anomalous requests to the ChatNVIDIA or reranking endpoints that contain suspicious local path string patterns in image parameters.