Banks generates meaningful LLM prompts using a simple template language. In versions prior to 2.4.4, all four media filters (image, audio, video, document) in b…

A path traversal vulnerability (CVE-2026-62663) exists in the 'banks' Python library, which is used to generate LLM prompts via a template language. In versions prior to 2.4.4, all four media filters (image, audio, video, document) accept untrusted user input as file paths without sanitization, canonicalization, or directory restriction. An attacker who controls template variables can use '../' sequences to read arbitrary files accessible to the Python process, including .env files, SSH keys, cloud credentials, /etc/passwd, and /etc/shadow. The file contents are returned base64-encoded within the rendered prompt output, making data exfiltration trivial. This is especially dangerous in applications that pass user-provided variables to banks before sending prompts to an LLM. The vulnerability has been patched in version 2.4.4.