Check files and images before the model reads them

How to do it

Accept only the types the feature needs, such as PNG, JPEG and PDF. Check the file's first bytes, not its name. Set limits below the provider's own, such as 5 MB per image and 20 pages per PDF. Reject larger files with a clear message. Rebuild every image as rebuild every uploaded image says, which drops EXIF location data before it leaves your servers. Scale images down to the size the task needs, because image tokens count toward the token budget. Text in an image or a PDF can carry hidden instructions, as the OWASP entry on prompt injection shows. Give the model no more power after reading a file than resist indirect prompt injection allows.

Examples

TypeScript example

const ALLOWED_TYPES = new Set(['image/png', 'image/jpeg', 'application/pdf']);
const MAX_IMAGE_BYTES = 5 * 1024 * 1024;
const MAX_PDF_PAGES = 20;

export async function checkModelFile(bytes: Buffer): Promise<string | null> {
  const type = (await fileTypeFromBuffer(bytes))?.mime;
  if (!type || !ALLOWED_TYPES.has(type)) return 'unsupported_type';
  if (type === 'application/pdf') return (await pdfPageCount(bytes)) > MAX_PDF_PAGES ? 'too_many_pages' : null;
  return bytes.length > MAX_IMAGE_BYTES ? 'too_large' : null;
}

Python example

ALLOWED_TYPES = {"image/png", "image/jpeg", "application/pdf"}
MAX_IMAGE_BYTES = 5 * 1024 * 1024
MAX_PDF_PAGES = 20

def check_model_file(data: bytes) -> str | None:
    kind = filetype.guess(data)
    if kind is None or kind.mime not in ALLOWED_TYPES:
        return "unsupported_type"
    if kind.mime == "application/pdf":
        return "too_many_pages" if len(PdfReader(io.BytesIO(data)).pages) > MAX_PDF_PAGES else None
    return "too_large" if len(data) > MAX_IMAGE_BYTES else None

How to check

Unit tests send a renamed program file, a 30-page PDF and a photo with GPS tags. The first two get a clear error with no model call. The photo reaches the fake provider with no EXIF data and its long side at or under the set limit.

Sources

• OWASP LLM01:2025 Prompt Injection

Related rules

• Test prompt injection in every file type the model reads in Security testing
• Rebuild every uploaded image in API design
• Keep every request inside a token budget in AI features
• Resist indirect prompt injection in Security testing