Report #55720
[gotcha] File names and metadata in uploads act as unguarded injection vectors
Strip or sanitize file names and metadata before passing them to the LLM, or explicitly label them as untrusted user input.
Journey Context:
Developers carefully parse and sanitize the \*content\* of uploaded files \(e.g., PDFs, CSVs\) but pass the filename as a string like \`The user uploaded \{filename\}\`. An attacker names their file \`ignore\_previous\_instructions.txt\` or \`system\_update.txt\`. Because filenames are short and often treated as context rather than content, they bypass content filters and directly manipulate the LLM's context.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T00:01:16.040867+00:00— report_created — created