howtosecure.ai
Implementation guide OWASP AISVS v1.01

Input Validation · Content & Policy Screening

Check images, audio and video for hidden attacks

AISVS 2.2.3 Level 2

Verify that non-text inputs (image/video/audio) are checked for adversarial perturbations, steganographic payloads, hidden or embedded content, or known attack patterns.

OWASP AISVS v1.01, requirement 2.2.3

In plain English

If you accept images, audio or video, check them for hidden instructions and manipulation before the model processes them. Text inside an image is read by a multimodal model just like typed text, and a file can carry content a person looking at it would never notice.

What good looks like

  • Text in images and speech in audio is extracted and run through the same injection and content checks as typed input.
  • Files are checked for known attack patterns and for data hidden in metadata or appended to the file.
  • Files are re-encoded before use, and metadata is stripped.
  • Suspicious files are blocked or quarantined, and the event is logged.

How teams usually do it

  • Optical character recognition (for example Tesseract or a cloud vision API) and speech-to-text ahead of the text checks.
  • Re-encoding, resizing or lightly compressing images and audio, which breaks many pixel-level perturbations and most steganography.
  • File type validation by content, plus a malware scan, before any model-specific check.

Suggested evidence

  • The processing steps for each media type you accept.
  • Test results using images with visible, faint and tiny embedded instructions.
  • Configuration showing metadata stripping and re-encoding.

Common mistakes

  • Screening typed text carefully and passing images straight to the model.
  • Running OCR at one resolution only. Text can be made readable to the model at a size or contrast your OCR misses.
  • Trusting the file extension.

Related requirements

2.2.4 Detect attacks that mix input types · 2.1.2 Catch encoding tricks hidden in input

Not sure which requirements apply to your system? Take the assessment, or browse every guide in this chapter.

Requirement text from the OWASP AI Security Verification Standard, licensed under CC BY-SA 4.0. This guidance is shared under the same licence.