Undocumented Behavior: Gemini 2.5 Flash Can Parse DOCX Files and Arbitrary Binary Data

Gemini API
,
1

Description

According to the official Vertex AI documentation for Gemini 2.5 Flash, the only supported MIME types for document inputs are:

Additionally, a Google representative stated in the forum that unsupported file types are extracted as pure text but without preserving structure.
[discuss.ai…google.dev]

Observed Behavior (Potential Issue)

While testing the API, I discovered:

  1. I can upload a DOCX file as raw binary data, and Gemini 2.5 Flash is able to answer questions about its content, even though DOCX is not listed as a supported MIME type.

  2. I tested further by uploading the raw binary of a compiled C program (.out / ELF).
    The model was able to extract and return the main() function signature, even though:

    • Executable binaries are not documented as supported inputs.

    • The API should normally reject unsupported MIME types, as seen in the StackOverflow report of .docx uploads causing a 400 Unsupported MIME type error.
      [stackoverflow.com]

This suggests the model backend attempts automatic text extraction from arbitrary binary data, which is not documented.

Why This May Be Important

  • The behavior is not described in the official documentation.

  • It differs from what users would expect based on the stated MIME limitations.

  • It may have security or privacy implications if the model automatically extracts strings from binary files.

  • It is unclear whether this is:

    • intended behavior,

    • an undocumented feature,

    • or a backend oversight.

Request

Could Google clarify:

  1. Whether extracting text from unsupported formats (including binary executables) is intended behavior?

  2. Whether this automatic extraction should be considered safe/production‑ready?

  3. Whether future documentation will explicitly address this behavior?

  4. Whether the API should reject non‑PDF, non‑text files more strictly?

Thank you!