Preventing prompt injection attacks through trusted file overlays

A system comprising:

• at least one computer processor;
• a file security manager, executing on the at least one computer processor;
• and
• a multi-modal large language model (LLM), executing on the at least one computer processor,
• wherein the file security manager is configured for:
• receiving a user-provided file as an input to the LLM,
• classifying, by a document classifier, the user-provided file to obtain a file type identifier of the user-provided file,
• retrieving, by the file security manager, a bounding box set of a trusted file corresponding to the file type identifier,
• masking, using a masking tool of the file security manager, the user-provided file with the bounding box set of the trusted file to obtain a secure file, wherein the masking the user-provided file comprises:
• overlaying a document image of the user-provided file with a plurality of bounding boxes in the bounding box set of the trusted file to obtain a plurality of content portions of the user-provided file,
• wherein the plurality of bounding boxes each encloses a content portion of the plurality of content portions of the user-provided file, and wherein the plurality of bounding boxes exclude a whitespace portion of the user-provided file,
• wherein respective content portions of the plurality of content portions correspond to respective bounding boxes of the bounding box set, and
• wherein the overlaying the user-provided file with the plurality of bounding boxes masks the user-provided file to retain the plurality of content portions of the user-provided file and blocks the whitespace portion of the user-provided file in order to obtain the secure file,
• initializing the secure file comprising a blank document image, and copying the plurality of content portions of the user-provided file, to the blank document image to obtain the secure file, and
• transmitting the secure file to the LLM, wherein the LLM is blocked by the secure file from processing a prompt in the whitespace portion of the user-provided file.

A method, comprising:

• obtaining a plurality of raw files from an external source, wherein a raw file comprises a document image;
• generating a plurality of corresponding bounding box sets for each respective raw file of the plurality of raw files, wherein the plurality of corresponding bounding box sets comprises a corresponding bounding box set comprising a bounding box, the bounding box including relative coordinates of a content portion of the document image;
• verifying the respective raw file of the plurality of raw files from the external source, based at least on metadata of the respective raw file, and the corresponding bounding box set of the respective raw file;
• generating corresponding file type identifiers for the respective raw file of the plurality of raw files, based at least on the metadata of the respective raw file;
• storing the plurality of raw files as trusted files, along with the corresponding file type identifiers and the corresponding bounding box sets in a trusted file repository;
• receiving, from a user application, a user-provided file, wherein the user-provided file comprises a second document image;
• classifying the user-provided file to obtain a file type identifier of the user-provided file;
• selecting the corresponding bounding box set of a trusted file corresponding to the file type identifier of the user-provided file from the trusted file repository;
• and
• overlaying the corresponding bounding box set of the trusted file on the second document image to obtain a plurality of content portions of the user-provided file, wherein the overlaying the user-provided file comprises:
• overlaying the document image of the user-provided file with a plurality of bounding boxes in the corresponding bounding box set of the trusted file to obtain the plurality of content portions of the user-provided file,
• wherein the plurality of bounding boxes each encloses the content portion of the plurality of content portions of the user-provided file, and wherein the plurality of bounding boxes exclude a whitespace portion of the user-provided file,
• wherein respective content portions of the plurality of content portions correspond to respective bounding boxes of the corresponding bounding box set, and
• wherein the overlaying the user-provided file with the plurality of bounding boxes masks the user-provided file to retain the plurality of content portions of the user-provided file and blocks the whitespace portion of the user-provided file in order to obtain a secure file,
• initializing the secure file comprising a blank document image, and
• copying the plurality of content portions of the user-provided file, to the blank document image to obtain the secure file,
• wherein, based on the overlaying, the corresponding bounding box set of the trusted file blocks a prompt in the whitespace portion of the second document image.